Accurate segmentation of the hippocampus and its subfields is critical for studying neurodegenerative diseases such as Alzheimer’s. These submillimeter structures require high-resolution MR images with strong signal- and contrast-to-noise ratios (SNR, CNR). These are difficult to achieve at 3T due to the hippocampus’s small size and deep brain location. A common approach for imaging the hippocampus is to use 2D T2-weighted (T2w) turbo-spin-echo (TSE) sequences, with very high in-plane resolutions and comparably thick slices. However, low SNRs and CNRs often compromise visibility of the fine internal hippocampal architecture. To address these limitations, within-subject averaging of multiple repeated acquisitions has been proposed to enhance image quality [1, 2]. When combined with robust preprocessing, such as N4 bias field correction and non-linear registration, this strategy can significantly boost anatomical fidelity. Here, we compare the generation of a high-SNR image using 25 repeated 3T T2w scans from two healthy subjects, aligned via a SyN-based ANTs pipeline. We compare simple averaging with a gradient-based method emphasizing structural boundaries, aiming to support improved visualization and ground truth annotation for deep learning-based hippocampal subfield segmentation.

Dataset. We acquired a unique dataset consisting of 25 repetitive T2w scans (3T Siemens Prismafit) of two healthy subjects. T2w scans were obtained using a 2D fast spin-echo with hyperechoes (48 slices, 0.47×0.47×1 mm³; 352×512×48 matrix) in an oblique coronal plane perpendicular to the long axes of the hippocampi. Image alignment. All T2-weighted volumes were aligned using a multi-stage registration pipeline implemented with the Advanced Normalization Tools (ANTs) software suite [3,4]. We applied symmetric normalization (SyN), a diffeomorphic model that generates smooth, invertible deformation fields suitable for high-resolution MRI registration [5]. To account for local intensity variations, common in T2-weighted imaging, the Mattes mutual information metric was used [6]. A four-level multi-resolution strategy (100, 70, 50, 20 iterations) with 32% voxel sampling per level was used to balance accuracy and efficiency [4,7]. The reference volume was selected based on the lowest Frobenius norm of intra-volume transformations between odd and even slices, indicating minimal inter-slice distortion and highest internal consistency. Prior to registration, all scans were N4 bias-corrected [8], padded by 20 mm in each spatial direction to reduce edge effects, and histogram-matched to the reference. Image Combination / Averaging. Following spatial alignment, which ensured a reliable basis for the subsequent voxel-wise high-SNR reconstruction, an average and gradient-based fusion approach was employed to synthesize a high-SNR T2-weighted volume from the aligned inputs. This method enhanced the representation of anatomical structures by prioritizing edges and tissue boundaries, which typically carry greater structural information. In particular, local gradient magnitudes were computed for each volume using the Gaussian gradient magnitude operator with a fixed standard deviation (σ = 1). Calculated gradients were then used to create weight maps. This process emphasizes smooth intensity transitions and edges while suppressing noise [9,10]. The final high-SNR volume was obtained by computing the weighted average voxel intensities. For comparison, also simple voxel-wise averaging of the aligned images was calculated.

To assess the quality of the resulting high-SNR images, SNR values were computed for the reference images and their corresponding batch-averaged reconstructions (N = 9, 16, 25). Quantitatively, the SNR of Subject 1 increased from 42.27 (original) to 59.99 using Gaussian-weighted averaging with a batch size of 25. For Subject 2, SNR improved from 35.56 to 54.44 under the same conditions. The complete set of SNR measurements across methods and batch sizes is summarized in Table 1. A qualitative comparison of the resulting images is shown in Figures 1 to 3.

Quantitative and qualitative assessments demonstrated consistent improvements in SNR across all reconstruction methods compared to the original T2-weighted MR images. Visual comparison (cp. Fig. 1 to Fig. 3) reveals enhanced anatomical detail and reduced noise in images reconstructed with Gaussian gradient weighting, particularly at higher batch sizes. Comparison of SNR values (Tab. 1) confirms that Gaussian weighting consistently outperforms simple averaging across all tested batch sizes.

This study demonstrates the benefit of generating high-SNR T2-weighted hippocampal images through averaging of repeated 3T acquisitions, with gradient-based Gaussian weighting consistently outperforming simple averaging. Resulting images offer improved anatomical fidelity, supporting more reliable downstream applications such as ground truth data for deep learning-based segmentation.

The monitoring diffuse lower-grade gliomas (DLGG) presents significant challenges due to the tumor's infiltrative nature and post-surgical brain remodeling, complicating MRI assessment. Traditionally, 2D tumor size measurements, as recommended by RANO criteria, have been used. Yet, they fail to capture subtle volumetric changes that are critical for tracking tumor progression. Volumetric assessment offers a more accurate estimation of tumor behavior but has so far been dependent on time consuming manual segmentation, rendering it impractical in routine clinical practice. Our previous proof-of-concept study demonstrated the feasibility of a nnU-Net model for DLGG segmentation, but clinical validation with larger, more heterogeneous datasets is essential. Indeed, existing literature reports encouraging results in low-grade glioma segmentation, but mainly on preoperative data, without incorporation longitudinal or post-surgical data.

This study was based on a cohort of 207 DLGG patients and 1971 MRI exams (Table 1) with longitudinal follow-up (9.55 exams ± 8.52). MRI imaging came from different MRI scanners (Siemens/GE/Philips) at different field strengths (1.5T/3T), including both 2D and 3D FLAIR acquisitions. The dataset, consisting of T1w and T2FLAIR (2D or 3D) data, was divided into a derivation set (N=1771) for model training and a validation set (N=200) for performance testing, while controlling for the 2D/3D FLAIR ratio in both sets (87.5%/12.5%). The nnU-Net model was trained using a 5-fold cross-validation approach. We evaluated the model's performance by comparing automated segmentations (AS) with manual segmentations (MS) performed by expert neuroradiologists. The primary evaluation metrics were the Dice Similarity Coefficient (DSC) and Intersection over Union (IoU), which measure segmentation overlap. Secondly, we derived tumor volume and mean tumor diameter (MTD) and compared AS with MS using Lin’s concordance correlation coefficient (CCC) and Bland-Altman tests. To further assess the model’s robustness, we analyzed its performance improvement across multiple training sets, gradually increasing the number of exams from 318 to 1971.

The nnU-Net model achieved a median DSC of 0.93 across both derivation and validation sets, with an IoU of 0.86. In the validation set, 64% of the cases showed a very good agreement (DSC ≥ 0.9) between AS and MS, and 31% had a good agreement (DSC between 0.7 and 0.9). Only 5% of cases showed unsatisfactory or poor results (Figure 1). Larger tumors were associated with higher DSC values (p<0.001). Tumor volume and MTD derived from AS showed near-perfect concordance with MS, with CCC values of 0.991 and 0.989, respectively for the validation cohort (Figure 2 and 3). Bland-Altman analysis showed a small underestimation by AS compared to MS, averaging 0.9 cm³ for tumor volume and 0.5 mm for MTD. The model’s performance improved as the number of training examples increased. With a smaller training set of 318 exams, the DSC was 0.82. Increasing the training data to 1009 exams improved the DSC to 0.89, and with 1771 exams, the DSC reached 0.93. However, adding more data beyond 1771 exams (up to 1971) did not yield further significant performance gains (p=0.84), indicating a plateau in the learning curve.

Our findings demonstrate that the nnU-Net model is a robust tool for automated DLGG segmentation, achieving a high level of accuracy that is comparable to manual expert segmentations. The model's performance, particularly in handling both 2D and 3D FLAIR images, shows its flexibility in real-world clinical settings. Importantly, the inclusion of longitudinal follow-up data and post-surgical cases with cavities makes this study more comprehensive compared to other studies, which focused solely on preoperative datasets. This broader scope allows for more accurate monitoring of DLGG progression over time and facilitates integration into clinical workflows. The minor underestimations observed in volume and MTD are unlikely to affect clinical decisions, suggesting that the model can be reliably used for patient follow-up. Furthermore, our analysis of the model’s performance across varying training set sizes confirms the importance of large, diverse datasets for improving deep learning model accuracy. The plateau observed in performance at 1771 exams suggests that the model has been sufficiently trained for this clinical context, although further external validation on datasets from different centers is recommended.

By significantly reducing the time required for segmentation while maintaining high accuracy, this model can enhance clinicians' ability to monitor tumor progression and assess treatment response in longitudinal follow-up. Future research should focus on external validation across diverse clinical environments to ensure generalizability and explore the model’s potential to provide more advanced clinical metrics, such as velocity of tumor expansion.

The Standard Model (SM) of white matter [1] describes the diffusion MRI (dMRI) signal as arising from the convolution of a fiber orientation distribution function (fODF) with a kernel comprising intra-axonal (sticks) and extra-axonal (zeppelin) water signal contributions. Fitting the SM to noisy dMRI data is challenging with its high-dimensional parameter space, potentially leading to low accuracy, precision, and degeneracy of estimates. Supervised deep learning has shown promise for fitting the SM to in vivo dMRI data [2-5], but it has potential drawbacks such as heavily relying on the choice of training data [6] and retraining for each acquisition scheme. Here, we implement Implicit Neural Representations (INRs) [7-9] for SM parameter estimation. INRs leverage spatial correlations across the brain to produce a continuous representation, in contrast to other methods that fit at voxel level. Furthermore, INRs are unsupervised, noise-robust, can be upsampled to higher resolutions, and can effectively include gradient non-uniformity correction in the fitting process. In this work, the INR method is compared against other machine learning methods and Nonlinear Least Squares (NLLS) fitting.

Ground Truth Generation: One MGH Connectome Dataset (subject 11) [10] was used for generating a ground truth. Acquisitions with Δ 19ms were selected. The SM was fitted with the SMI toolbox [2] to generate a set of realistic SM parameters for axon signal fraction f, intra-axonal diffusivity Di, extra-axonal parallel diffusivity De , extra-axonal perpendicular diffusivity Dp. An anisotropic diffusion filter was applied to generate spatially smooth ground truth maps. The Spherical Harmonics (SH) coefficients plm of the FOD were calculated using Multi-Shell Multi-Tissue Constrained Spherical Deconvolution (MSMT CSD) [11]. The simulated signals corresponding to these parameters were calculated from the SM signal equation with an optimized acquisition protocol [2]: b-values [0, 1000, 2000, 8000, 5000, 2000]mm/s2 , number of directions [6, 20, 40, 40, 35, 15] and B-tensor shape [1, 1, 1, 1, 0.8, 0]. A maximum SH order of lmax = 2 was used and Gaussian noise was added (SNR 50). A signal mask was applied based on white matter segmentation [12]. Fitting methods: The SM was implemented as INR (1) in Pytorch (Figure 1) and compared to three other SM fitting methods: 2) supervised machine learning using the SMI toolbox [2] with standard settings; 3) supervised deep learning as implemented in [3] trained on uniformly distributed parameters, 5e5 samples, 75%/25% training/validation split, Gaussian noise with SNR=50; 4) NLLS implemented in the MATLAB optimization toolbox using Levenberg-Marquardt algorithm . Method 1 was trained on a NVIDIA Titan X GPU for 150 epochs using a mean-squared-error loss. Pearson correlation coefficient ρ and Root-Mean-Squared-Error (RMSE) were used for evaluation. Gradient non-uniformity correction: One healthy volunteer was scanned on a 3T, 300 mT/m Connectom scanner (Siemens Healthineers) with b-values up to 4000 mm/s2 and B-tensor shape [1, 0.5, -0.5, 0]. SM parameter estimation was performed with the INR without and with correction for gradient-nonuniformities. In the second approach, spatially varying, scanner-specific gradient deviations were used to compute a B-tensor for every voxel [13]. This corrected B-tensor was then used in the forward signal prediction (Figure 1c).

Fig. 2a shows scatter density plots of the SM parameters for fitting approaches 1 to 4 . RMSE and ρ show superior performance of the INR model for all parameters, most significant for Di and De . Fig. 2b presents the resulting parameter maps across the brain, with INR representing the white matter structure smoothly, while the other methods show spatial irregularity. Fig. 3 highlights the impact of incorporating gradient non-uniformity correction on parameter estimation, with a notable reduction in De values observed when the correction is applied. Training on the simulated data took 308 seconds, whereas training with gradient non-uniformity correction required 534 seconds.

Validation on simulated data demonstrated that the INR method more effectively mitigates noise by leveraging spatial correlations compared to voxel-wise fitting methods. Gradient non-uniformity correction is cumbersome in supervised machine learning as training data would have to be generated capturing all B-tensor variations. INRs offer straightforward gradient-nonuniformity correction while being independent from training data. The INR method shows reasonable interference times, much faster than NLLS.

INR-based approach enables accurate and robust estimation of white matter microstructure, outperforming traditional methods by leveraging spatial continuity, incorporating gradient non-uniformity correction, and eliminating the need for training data.

Optimizing diffusion MRI (dMRI) acquisition protocols in clinical settings is essential for improving image quality, reducing scan times, and enhancing sensitivity to tissue microstructure. Traditional optimization methods have mainly focused on minimizing the Cramer-Rao Lower Bound (CRLB)[1], optimizing angular coverage of b-shells[2], or using data interpolation[3] to improve model parameter estimates during MR acquisition design. However, machine learning and deep learning methods, such as TADRED (Task-Driven Experimental Design)[4], or physics-informed networks[5] can generalize more flexibly to different dMRI models. TADRED is a deep learning framework that combines a subsampling network, which identifies the most informative measurements, with a task network that performs the task. This subsampling process efficiently identifies the most informative subsets of data while training a high-performing task-specific model. TADRED optimizes acquisition protocols by progressively reducing the sample set, enhancing overall efficiency. In this study, we demonstrate the TADRED-based subsampling approach against a naïve subsampling approach that aims to maximise angular coverage within each shell.

Data Acquisition, Processing and Fitting: Multi-shell dMRI data[6] were acquired on an ultra-strong gradient system (TE = 80 ms, TR = 5 s) with 1.8 mm slice thickness, 66 slices, and a 120 × 120 mm² FOV. The dataset included seven b-shells up to b = 6 ms/μm² and 266 directions (13 b=0). Gradient spacing (Δ) and duration (δ) were 24 ms and 7 ms. Direction counts scaled with b-values, with interleaved b=0 images to aid correction. Gradient spacing (Δ) and duration (δ) were 24 ms and 7 ms, respectively. Number of directions scaled with b-values, and the b=0 images were interleaved through the acquisition for improved correction. Data were denoised with MP-PCA[7], [8], [9], [10], corrected for drift, outliers (SOLID[11]), and distortions from susceptibility, motion, and eddy currents were corrected using FSL’s topup[12], [13] and eddy[14]. Gradient non-linearity was corrected using MATLAB, and Gibbs ringing was removed using MRtrix3[15]. For the TADRED, SANDI[16] model parameter maps were generated from the full dataset with a random forest regression as base parameter maps, using the SANDI MATLAB Toolbox[17]. Data Subsampling and Analysis: Two subsampling approaches were used to optimally subsample the dMRI data: uniform subsampling and TADRED framework[4]. In the first approach, data were reduced uniformly by 10%, 30% and 50% from each b-shell, ensuring that the remaining directions are still distributed uniformly on the unit sphere. For the latter approach, the dMRI data, the generated SANDI maps and acquisitions parameters were concatenated and provided as an input to TADRED framework. The TADRED subsampling and task networks were trained on these SANDI maps using 84% of the data for training and 8% each for validation and testing. Finally, TADRED generated desired subsampled datasets without any constraints from the fully sampled dMRI data. SANDI parameters maps were generated for uniformly and TADRED-subsampled datasets in the same way as the full dataset, which were compared to ‘gold-standard’ maps to evaluate the accuracy and effectiveness of each subsampling method.

Fig. 1 shows the subsampled diffusion directions across several b-shells for both uniform and TADRED approaches. Notably, TADRED removed more directions at the intermediate b-shell (b = 2.4 ms/μm²). Fig. 2 displays the TADRED-predicted SANDI maps alongside ‘ground-truth’ maps for subsampling rates of 10% (A) and 50% (B). Difference maps highlight parameter discrepancies between TADRED outputs and the ground truth. Fig. 3 compares SANDI model parameter maps derived from the uniform and TADRED-subsampled datasets, illustrating variations between the two approaches. Fig. 4 presents the mean squared error (MSE) and mean absolute error (MAE) of the SANDI parameters, averaged across all brain volumes.

TADRED test results show less than 20% error within brain tissue, with higher errors primarily observed in cerebrospinal fluid regions. In general, TADRED demonstrates lower errors in the difference maps of SANDI parameters across all subsampling rates compared to uniform subsampling (Fig. 3). However, at the 50% subsampling level, errors in fneurite and Rsoma are higher than those from uniform subsampling (Fig. 4); however, the higher error in the voxels originates in the CSF. In this study, we only utilised the TADRED subsampling network to focus on its protocol optimisation capabilities, applying the same model fitting method to both optimisation approaches. Future work will incorporate the TADRED task network to directly estimate the SANDI parameters.

TADRED outperformed uniform subsampling in estimating SANDI parameters, showing lower error rates and promising potential for efficient dMRI protocol optimization in future clinical and research applications.

Saturation transfer (ST) MRI has shown promise in various molecular imaging tasks [1–5]. However, its traditional contrast-weighted form is affected by multiple confounding tissue and pulse-sequence parameters [2,6]. Quantitative ST techniques, such as QUESP, yield improved specificity, yet assume steady-state conditions that require a relatively long acquisition [6,7]. ST Magnetic resonance fingerprinting (MRF) overcomes these challenges by matching transient experimental signals to a precomputed dictionary [8–13], predominantly under preclinical settings or continuous wave (CW) acquisition. However, to render ST MRF a clinically viable method, the acquisition protocol must be shortened, optimized for clinical scanners, and accommodate rapid pulsed wave (PW) saturation [12]. Deep learning-based MRF pipelines [14, 15] can efficiently generate optimized acquisition protocols. Yet, to date, they were only compatible with CW acquisition, due to their inherent reliance on analytical solutions of the Bloch-McConnell (BM) equations. Here, we developed a fully differentiable framework for Automatic end-to-end discovery of Pulsed ST MRF sequences (AutoPulST).

A differentiable ST computational graph was implemented in JAX [16], unlocking the ability to rapidly extract the numerical solution of the BM equations. The solver was wrapped within an automatic acquisition protocol discovery framework [14], where a multi-layer perceptron quantification network is jointly trained (Fig. 1). The pipeline was implemented for discovering protocols composed of 4, 8, or 16 raw MRF images (representing a 1.9-7.5 fold acceleration). Before each optimization attempt, a baseline reference protocol was randomly generated. Imaging was performed at 7T (Terra, Siemens Healthineers) and 3T (Prisma, Siemens Healthineers). Pulsed ST-MRF sequences were automatically generated and implemented for two imaging scenarios: L-arginine phantom imaging (25-100 mM, pH 4–6, chemical shift = 3 ppm, room temperature, 7T) and semi-solid MT brain imaging of healthy volunteers (3T, following IRB approval and informed consent). The backbone pulse sequence used a spin lock saturation train (13x100 ms, 50% duty-cycle), which varied the saturation pulse power and frequency offset. A 3D centric reordered EPI (3T) or GRE (7T) readout module was used [17, 18]. For comparison, a previously established MRF protocol was implemented, which acquires 30 raw images within 143.4 s (3T) or 325.6 s (7T) [19].

In vitro imaging at 7T: the quantitative parameter maps obtained using AutoPulST demonstrated improved SNR compared to the random reference protocols and were visually similar to the gold standard (Fig. 2A). The absolute percent errors obtained by the full AutoPulST pipeline (namely, simultaneous optimization of the acquisition and quantification network) were more than 50% lower than those obtained by the reference random scan (p<0.001. Fig. 2B). Human brain imaging at 3T: A clear visual improvement was obtained using AutoPulST generated protocols (Fig. 3) compared to the random acquisition protocols. The AutoPulST optimized protocols provided statistically significant quantification improvement (p<0.001 across 12 optimization attempts, Fig. 4) compared to the random alternatives and were acquired within merely 19.1, 38.2, and 76.5 s.

In this study, only two acquisition parameters (B1 and Δωrf) were optimized. Nevertheless, we expect that incorporating additional parameters (e.g., saturation pulse time, recovery time, etc.) could further improve the quantification ability.

We demonstrate a differentiable, end-to-end optimization framework for rapid pulsed ST-MRF. We expect this approach to play a key part in the translation efforts of quantitative and rapid ST imaging.

Intravoxel Incoherent Motion (IVIM) is a diffusion-weighted (DW) MRI technique that models signal decay from tissue diffusion and capillary blood flows [1]. Deep Neural Networks (DNNs) have emerged as a powerful alternative to traditional fitting methods, offering fast and robust inference with respect to conventional non-linear least squares (LSQ) and Bayesian approaches [2], [3]. DNNs typically provide only point estimates without uncertainty, which can help in understanding the impact of noise and guiding experimental design [4]. This study, funded by a PRIN 2022 project [5], introduces a set of methods and metrics, based on the use of Mixture Density Networks (MDNs) [6] with Deep Ensembles (DEs) [7] enabling, in a supervised voxel-wise setting, the estimation of both aleatoric (AU) and epistemic uncertainty (EU) and model calibration in IVIM.

We trained our networks using simulated images based on the Shepp-Logan phantom in MATLAB, covering a wide range of IVIM parameters: D (0-0.003 mm²/s), f (0-0.4), and D* (0.003-0.2 mm²/s). The images were computed at the same b-values as those used in our in vivo acquisitions, employing the model equation: S(b) =S₀[f · e–b·D* + (1 – f) · e–b·D ] Rician noise was incorporated with SNRs of 25, 50 and 100, for a total number of 6000 phantoms. We used one in-vivo dataset for testing the networks in a real scenario: mouse brain data (n=6) acquired using a 7T Bruker Biospec using EPI-DWI with 14 b-values (0-1000 s/mm²). Acquisition parameters: TR = 3 s, TE = 61 ms, 2 segments, 2 repetitions, FOV =15 × 15 mm, MTX = 76 × 76, slice thickness = 0.5 mm, δ = 5.8 ms, ∆ = 50 ms. Architecture We implemented three voxel-wise models: an MLP, an MDN with either three or one Gaussian. All models used two hidden layers with ELU activations and a number of neurons equal to the number of b-values. Inputs were IVIM signals at each b-value, normalized by the signal at b = 15 (excluding b = 0 due to artifacts in-vivo). Sigmoid activations were applied to the output and scaled to the physiological ranges of the IVIM parameters. To capture overall uncertainty, we trained an ensemble (M=5) of MDNs with different random initializations. For each MDN in the ensemble, the predictive mean and variance were computed as a weighted average of its Gaussian components. Then, AU was computed as the average predictive variance across ensemble members, while EU was estimated as the variance of the predictive means of each MDN [7]. Experimental The dataset was split into 70% training, 20% validation, and 10% test sets, maintaining the SNR distribution across subsets. Adam optimizer (learning rate 0.0001, batch size 32) for up to 600 epochs, with early stopping (patience of 35) based on validation loss was used. The MLP used Mean Squared Error loss, while probabilistic models optimized the negative log-likelihood [8], [9]. For simulations, accuracy was assessed using the Median Absolute Error (MdAE), the Relative Median Bias (MdB) and Robust Coefficient of Variation (RCV) [10]. To quantitatively evaluate the quality of uncertainty estimates, we employed calibration reliability diagrams, and the Continuous Ranked Probability Score (CRPS) [11], with lower CRPS indicating better uncertainty estimates.

Tab.1 shows the accuracy metrics on the simulation test set across different SNR levels. MLP, MDN, and Gaussian models perform similarly and outperform the Bayesian fitting method. Fig.1 displays an MDN prediction on a Shepp-Logan phantom at SNR 50, where AU dominates due to noise (mean AU: 4.6×10⁻⁵ mm²/s for D, 0.02 for f and 0.02 for D*), while EU remains low (mean EU: 4.7×10⁻6 mm²/s for D, 0.001 for f and 0.001 for D*). The reliability diagrams of MDN and Gaussian models are reported in Fig.2, showing better calibration for MDN, as evidenced by a smaller miscalibration area. CRPS values also favor MDN, with slightly lower scores: D (5×10⁻⁵ vs 6×10⁻⁵ mm²/s), f (0.019 vs 0.021), and D* (0.021 vs 0.022 mm²/s). Fig.3 shows an MDN prediction on an in-vivo slice, where EU - especially for f - is notably higher (around 0.02) than in the simulated case (Fig.2), reflecting the model’s increased uncertainty on real data.

We adopt, for the first time in IVIM, probabilistic models like MDNs with DEs to estimate both AU and EU. EU, even though much lower than AU in our study, is often overlooked in quantitative MRI but is crucial for detecting out-of-distribution data. The elevated EPU for f in Fig. 3 may suggest a mismatch between the broad simulation range (0–0.4) and in-vivo values. Quantitative metrics for calibration and CRPS allowed to assess the better reliability of uncertainty estimation of MDN.

Our results demonstrate the value of probabilistic modeling for IVIM parameter estimation and uncertainty quantification. Future work will focus on incorporating spatial context via convolutional architectures (e.g., CNNs) and exploring alternative probabilistic approaches such as Normalizing Flows.

Glioblastoma (GBM), which receives a grade IV classification from the World Health Organization (WHO), is the most prevalent malignant brain tumor in adults [1]. Differentiating between TrueProgression (TP) and Pseudoprogression (PsP) remains a diagnostic challenge, as both appear to be similar on structural MRI sequences like T1 post-contrast (T1pc) and T2 FLAIR. Misclassification of TP or PsP may result in unnecessary interventions or delayed treatment actions, which diminishes the overall patient outcome [2]. Dynamic contrast enhanced MRI (DCE-MRI) based pharmacokinetic (PK) modeling adds important information about tumor vascularity and perfusion. This study employs a parsimonious DCE-MRI model to derive biologically relevant vascular parameters (Ktrans, Ve, and Vp), which are further analyzed through radiomic feature extraction using PyRadiomics to enhance the detection of subtle patterns[3]. Machine learning models were assessed based on their capabilities to classify TP vs. PsP. Lasso, elastic net, and ridge regression were used for feature selection to enhance model performance.

The study featured a private DCE-MRI dataset from the University of Pennsylvania, USA and shared with TCG CREST (data sharing ID: RIS76150), containing 57 GBM cases with 38 TP and 19 PsP. Prior to image processing, a parsimonious modeling approach was used to fit DCE-MRI data and extract pharmacokinetic parameters. Tumors were segmented into enhancing, non-enhancing, and edema regions using the nnU-Net algorithm [4]. The dataset was split into 70% training and 30% testing sets, and SMOTE[5] was applied to correct the 2:1 class imbalance between TP and PsP. Five pharmacokinetic models (Non-linear Tofts, Extended Tofts, Shutter-Speed, 2CXM, and 3S2X) were voxel-wise fitted to generate R² and AIC maps[6].The model with the lowest AIC at each voxel was selected as the best fit. A parsimonious model was then derived by aggregating these voxel-wise best fits to estimate Ktrans (volume transfer constant), Ve (extravascular extracellular volume fraction), and Vp (plasma volume fraction) [7] as shown in Fig 1. Total of 280 radiomic features (140 for enhancing and 140 for non-enhancing regions) were extracted from the pharmacokinetic maps to quantify tumor heterogeneity. Four machine learning classifiers were trained on the selected features: Random Forest (RF), Support Vector Classifier (SVC), Logistic Regression, and XGBoost. Each model were trained and evaluated with and without feature selection to assess the impact of dimensionality reduction on classification performance. The classification performance was assessed using Accuracy and F1 Score. The pharmacokinetic parameters were analyzed separately to determine their predictive power. To improve classification performance, we applied three different feature selection techniques—LASSO, Ridge, and Elastic Net—to identify the most relevant features.

Fig. 2 presents the best-performing models for each feature selection method and pharmacokinetic parameter, along with their corresponding accuracy and F1-score. The highest-performing model in each category is highlighted in bold. ROC curve for all the models after feature selection using Elastic Net is shown in Fig 3 for the Ktrans parameter. For each estimated pharmacokinetic parameter, the important features after elastic net and gini importance (Fig 4) included: original shape - least axis length, GLSZM size zone percentage, original shape - sphericity, first order -variance, root mean squared, total energy, range.

Random Forest remained a strong performing model, consistently performing well due to its abilities to address underlying non-linearity and feature interactions. The ensemble approach of Random Forest also assisted in minimizing overfitting while working with a small dataset. In a similar manner, SVC (with RBF kernel) produced strong results, especially for Ve, likely due to the manner in which it can build non-linear decision boundaries in high dimensional feature spaces. Elastic Net addressed the challenges of overfitting with the balance of L1 and L2 regularization while selecting the most informative features. This study supports that the maximum classification performance occurred when Elastic Net, with Random Forest, was applied to Ktrans, achieving F1-score of 92.86% with accuracy of 88.89%.

The present study helps advance the use of machine learning applications in medical imaging by considering pharmacokinetic parameters during classification of the tumor type. An important limitation of this study is the small sample size which limited appropriate validation and increased the risk of overfitting. Future studies can explore multiparametric approach combining multiple features may further improve the robustness of tumor classification models. Additionally, as dataset sizes increase, deep learning architectures can increase the classification accuracy.

Arterial Spin Labeling (ASL) is a non-invasive MRI method for quantifying cerebral perfusion via magnetically labeled blood water. Time-encoded pseudo-continuous ASL (te-pCASL) with Hadamard encoding enhances signal efficiency and enables simultaneous cerebral blood flow (CBF) and arterial transit time (ATT) estimation [1]. The Buxton model describes the ASL signal evolution and is widely used for perfusion quantification [2], though nonlinear least squares (NLLS) methods often yield unstable parameter maps under low signal-to-noise conditions due to sensitivity to noise and initialization. Physics-Informed Neural Networks (PINNs) offer a robust alternative by embedding governing physical models into the learning process, improving resilience to noise and data sparsity [3]. Their applications in biomedical imaging are expanding, including cardiovascular modeling [4], perfusion CT [5], and neonatal ASL [6]. Building on [6], we extend the PINN framework to te-pCASL and assess its performance in a synthetic simulation study.

We designed a two-stage PINN framework trained on synthetic te-pCASL data generated using spatially heterogeneous CBF, ATT, and equilibrium magnetization (M₀) derived from Boston ASL Template and Simulator [7]. A central slice was extracted and resampled to match the voxel size of real ASL datasets (3x3x7 mm3). The actual signals were simulated using a custom implementation of the Buxton model, accounting for sub-bolus-specific labeling durations (Figure 1). Gaussian noise was added to yield low temporal signal-to-noise ratio conditions (tSNR = 1.5 and 0.5), following methods from prior studies [8]. The model comprised two coupled SIREN-based Multi-Layer Perceptrons (MLPs): a data-fitting network (7 layers, 256 units) mapping spatial coordinates (x, y) and time to ASL signal, and a physics-based network (4 layers, 128 units) mapping (x, y) to CBF and ATT values (Figure 2). To consider the dependence of bolus duration on acquisition time in absence of an analytic expression, a fifth-order spline was used to model this relationship, and its derivative was included in the Buxton model derivative during training to better reflect time-encoding dynamics. Training minimized a hybrid loss combining mean squared error (MSE) between predicted and simulated signals, and MSE between time derivatives of the data-fitting network and those obtained from the Buxton model derivative using predicted parameters. Optimization used Adam (learning rate=0.0001) for 30,000 epochs with a batch size of 150. At inference, only the physics-based network was used for parameter estimation. As a baseline, we used a regularized NLLS estimator with soft-L1 loss, physiological constraints, and optimization via Dogbox. It minimized residuals between signals and Buxton model predictions, with mild regularization on ATT deviations from initialization. Performance was evaluated using Structural Similarity Index Measure (SSIM), Pearson Correlation Coefficient (r), and Percentage Relative Error (PRE), along with visual inspection for qualitative comparison.

At tSNR = 1.5, both methods performed similarly for CBF. However, PINN significantly outperformed NLLS in ATT estimation, a parameter more sensitive to noise (Table 1). At tSNR = 0.5, PINN’s advantage was even more pronounced, achieving higher SSIM and r values for both parameters. Qualitatively, PINN produced smooth, physiologically explainable maps (Figure 3). While ATT overestimation persisted, PINN better preserved spatial structure. Incorporating the spline correction reduced ATT bias, indicating better modeling of bolus-specific dynamics, though a residual bias remained. NLLS maps, in contrast, showed higher noise and voxel-wise variability, particularly for ATT.

Embedding the Buxton model in a PINN framework enhances robustness in perfusion quantification under noise. In particular, ATT estimation benefits significantly, addressing limitations of conventional methods. The spline-based correction partly mitigated ATT bias, supporting the hypothesis that bolus duration variability influences estimation accuracy. Nevertheless, the remaining bias suggests additional sources of mismatch, possibly due to simplifications in the kinetic model or training dynamics. Future work will explore further physical model refinements and apply the framework to real ASL datasets.

In conclusion, we propose a PINN-based framework for ASL quantification, integrating the Buxton model as a physical constraint. Compared to NLLS, our method improves ATT estimation and noise robustness, while maintaining physiologically plausible spatial patterns. This work lays the foundation for extension to 3D and clinical ASL studies.

Convolutional neural networks (CNNs) have shown strong performance in classifying Alzheimer’s disease (AD) from T1-weighted (T1w) MRI [1], yet their decision-making remains largely opaque [2]. Ensuring decisions are not driven by spurious features is critical for clinical applications. Prior work [3] indicated that CNNs favor skull-stripped over full images for T1-based AD classification. However, how learned features are spatially organized under different preprocessing pipelines remains unstudied. Spectral clustering (SC) [4] can group similar heatmaps and reveal dominant decision patterns beyond individual cases. This study applies Spectral Relevance Analysis (SpRAy) [5] to Layer-wise Relevance Propagation (LRP) [6] heatmaps in a CNN trained on T1w MRI. We assess whether relevance patterns vary across preprocessing methods and identify brain regions that consistently drive classification, allowing insights into model behavior and potential preprocessing biases.

Participants. From the ADNI database (https://adni.loni.usc.edu), we selected 1,042 AD and 2,227 normal control (NC) T1w MRIs acquired sagittally at 3T with consistent resolution. A balanced dataset was constructed with 990 images each from 159 AD and 201 NC participants (propensity-score matched by age and sex) [7]. Scans were split into training (70%), validation (15%), and test (15%) sets with subject-level separation. Preprocessing. Images were reoriented, bias-corrected [8], non-linearly registered to MNI space [9], and normalized to the white matter histogram peak. Skull-stripping was performed using FSL-SIENAX [10]. Binarization thresholds of 13.75%, 27.5%, and 41.25% of white matter peak intensity were applied to generate texture-masked images, resulting in eight preprocessing conditions (aligned/skull-stripped × 4 thresholds, see Fig. 1). CNN Architecture. We adapted a 3D CNN from [1] with reduced complexity to avoid overfitting. For a schematic overview see Fig. 2. Training. Each configuration was trained for 30 epochs (batch size 20) using Adam [11]. Ten random data samplings per configuration were used [12], with three models trained per sample (total: 30 per setup) [13]. The best validation-performing model per setup was used for further analysis. Heatmap Generation and Spectral Clustering. LRP (α=1.0, β=0.0) [6] generated voxel-wise relevance maps. Mean heatmaps were computed and thresholded to the top 40% of relevance for inspection. SpRAy involved: 1. Generating LRP heatmaps, 2. Downsampling to 2 mm resolution, 3. Spectral clustering of relevance maps, 4. Identifying meaningful eigenvalue gaps, 5. Visualizing clusters via t-distributed stochastic neighbor embedding (t-SNE) [14].

Skull-stripped data without binarization (A2) achieved the highest accuracy (81.6%). Other configurations, including 27.5% binarized skull-stripped images (C2) and 41.25% binarized aligned images (D1), performed comparably. For full statistics, refer to [3]. t-SNE visualization (Fig. 3) showed that only model D1 produced heatmap clusters aligning with subject groups (AD vs. NC). This suggests that omitting skull-stripping and using 41.25%-threshold binarization may lead to more positionally distinctive relevance distributions. Clustered mean heatmaps (Fig. 4) showed D1’s Group 2 highlighting the left insular cortex -a known AD-affected area- while Group 3 included skull regions, pointing to potential reliance on non-brain features. Models A2 and C2 showed less separation and less anatomically interpretable relevance.

This study extends prior findings on CNN reliance on volumetric cues [3] by showing how preprocessing impacts relevance patterns. SC enabled group-level analysis of LRP heatmaps, identifying systematic rather than instance-specific decision strategies. Only D1, the model trained on non-skull-stripped and 41.25% binarized images, revealed clustering aligned with AD vs. NC labels. Its mean heatmaps featured clearer anatomical relevance (e.g., left insular cortex), but also highlighted skull regions, suggesting possible biases. This reveals a trade-off: less aggressive preprocessing may preserve informative structure but also introduce non-brain artifacts. Importantly, similar classification performance across all configurations obscures these interpretability differences, underscoring the need for explainability techniques in model evaluation [15]. Preprocessing choices, even minor ones, can substantially alter feature attribution.

We applied SC to LRP heatmaps from CNNs trained on differently preprocessed T1w MRIs for AD classification. Despite similar accuracies, relevance patterns varied markedly, with the 41.25% binarized model without skull-stripping (D1) showing clearer separation between AD and NC. Our findings stress that preprocessing has a strong impact on interpretability, not just performance. Future work should explore the effects of preprocessing and the integration of quantitative MRI features to improve model transparency and clinical utility.

While deep learning (DL) methods in MR image reconstruction are becoming state-of-the-art, the possibility of DL models altering the appearance of pathology still raises concerns, especially when trained solely with healthy subjects and/or when employed on patients’ datasets acquired with higher undersampling factors [1]. This issue is particularly relevant and not sufficiently examined in the case of clinical acquisitions of the brain with submillimeter resolutions, where undersampling is highly desirable to reduce acquisition time. In this study, a DL-based reconstruction was investigated for the acceleration of patients’ clinical scans acquired with 0.6 mm isotropic resolution at 7T. Images were retrospectively undersampled to test different accelerations, and image quality and pathology conspicuity were compared against conventional compressed sensing (CS) reconstruction.

Study Population and MR Acquisition Six patients were enrolled in this preliminary analysis: three patients with different brain tumors and three patients with suspected multiple sclerosis (MS). All patients were scanned at 7T (MAGNETOM Terra, Siemens Healthineers, Forchheim, Germany) using an 1Tx/32Rx RF head coil (Nova Medical, Wilmington, USA). A 3D MP2RAGE research application sequence (resolution = 0.6×0.6×0.6 mm3, matrix size = 384x256x384, TR/TI1/TI2 = 6000/800/2700 ms, TE/echo-spacing = 2.06/6.2 ms, acquisition time = 7:40 min) with a 4x accelerated spiral phyllotaxis sampling was acquired [2–4]. The work was approved by the local ethics committee (Kantonale Ethikkommission für die Forschung Bern, Switzerland, 2020–02902, 2022–00720, 2024-00788). Prior to each examination, written informed consent was obtained. Image Reconstruction MP2RAGE acquisitions were also retrospectively undersampled to simulate acceleration factors of R=8 (3:50 min) and R=10 (3:04 min). All images were reconstructed using both a conventional CS reconstruction [5,6] and a 3D DL-based method [7]. The DL technique reconstructs images from undersampled k-space data and coil sensitivity maps using six iterations of data consistency updates and neural network evaluation. The network was initially trained with 5000 pairs from fully-sampled 3D datasets of healthy subjects scanned at 1.5 and 3T, followed by fine-tuning in a self-supervised manner using 1000 pairs of undersampled 3D datasets collected at 7T [8,9] (all MAGNETOM scanners, Siemens Healthineers, Forchheim, Germany). Image Analysis A visual inspection and a quantitative quality assessment of retrospectively reconstructed images was conducted by computing the structural similarity index measure (SSIM) and the peak signal-to-noise ratio (PSNR) considering the current clinical protocol (R=4, CS reconstruction) as reference.

Reconstructed images for patients with suspected dysembryoplastic neuroepithelial and epidermoid tumor are shown in Figure 1-2. Overall, DL reconstructions exhibit higher SNR than CS reconstructions while preserving tissue contrast. For the highest tested acceleration (R=10), better (Figure 1) or similar (Figure 2) conspicuity of tumor boundaries and higher SNR was observed in the DL reconstruction compared to CS. Images from two of the MS patients are reported in Figure 3-4. Lesions identified with acceleration R=4 could be visualized at higher acceleration rates. Lesion conspicuity was found to be similar or better in DL reconstructions in comparison to CS. None of the images reconstructed with the DL reconstruction exhibit hallucinations or misleading artifacts. For R=8, CS reconstruction achieved an SSIM of 0.99±0.01 and a PSNR of 37.3±0.3 dB, while DL reconstruction yielded an SSIM of 0.99±0.01 and a PSNR of 38.1±0.5 dB. At R=10, CS reached an SSIM of 0.98±0.01 and a PSNR of 35.6±0.2 dB, whereas DL reconstruction achieved an SSIM of 0.99±0.01 and a PSNR of 37±0.4 dB.

In this preliminary study, we evaluated the performance of a DL-based reconstruction method trained solely on healthy subjects to reconstruct images of patients with different brain abnormalities. No hallucinations or misleading artifacts resulting from the DL reconstructions were observed. Results were compared to a CS reconstruction, showing that the DL reconstruction exhibits higher SNR, PSNR and equivalent or improved conspicuity in the observed pathologies. These findings suggest robustness of the DL approach across different pathological presentations. The work conducted in this study is planned to be extended to a larger cohort of patients including radiologist rating of the reconstructed images, while being blind to the acceleration factor or the reconstruction technique, to establish the clinical value of DL reconstruction more comprehensively for 7T neuroimaging.

This preliminary study demonstrates the quality and reliability of DL-based MRI reconstruction in the presence of pathology, even at high acceleration rates. In the future, we will validate these results in a larger cohort of patients.

Gliomas are primary brain tumors characterized by marked biological heterogeneity, affecting prognosis and treatment strategies [1,2]. Traditional histopathological grading, though informative, is invasive and limited by sampling bias [3,4]. Non-invasive radiomic analysis using MRI has shown promise in assessing tumor grade [5,6]. However, most machine learning (ML) studies frame glioma grading as a classification task, potentially overlooking subtle intra-class variability [7]. In this study, we propose a regression-based ML approach to predict glioma grade as a continuous variable using radiomic features extracted from T1-weighted and diffusion tensor imaging (DTI) to enhance diagnostic precision and interpretability [8,9].

We analyzed MRI data from 36 glioma patients (58.3% low-grade gliomas [LGG], 41.7% high-grade gliomas [HGG]) acquired on a 1.5T scanner. Radiomic features were extracted from normalized T1-weighted images and diffusion maps, including axial diffusivity (AD), radial diffusivity (RD), mean diffusivity (MD), and fractional anisotropy (FA) [10,11]. Feature extraction included intensity, shape, and multiple texture matrices (GLCM, GLRLM, GLSZM, NGTDM, GLDM) using PyRadiomics [12]. Dimensionality reduction was performed using Sequential Feature Selection (SFS) with a Random Forest regressor [13]. A total of 15 ML regression models, including XGBoost, CatBoost, SVM, and neural networks, were trained and optimized via stratified 5-fold cross-validation. Model performance was evaluated using MSE, MAE, and R² scores. SHAP analysis was applied to interpret the contribution of individual features [14], as showed in Figure 1.

Among all models, XGBoost achieved the best performance with an MSE of 0.0346 ± 0.0137, MAE of 0.1168 ± 0.0256, and R² of 0.5140 ± 0.3579 (Figure 2). Feature selection identified six key predictors, with texture features from T1-weighted images (e.g., GLRLM short run low gray level emphasis, GLCM contrast) being the most influential (Figure 3). Diffusion-derived features, particularly robust mean absolute deviation from AD maps, added complementary information. SHAP analysis highlighted how selected features differentially contributed to grade predictions across LGG and HGG samples, with interpretable patterns aligning with tumor microstructural complexity (Figure 4) [15,16].

Modeling glioma grade as a continuous variable allowed for finer granularity in characterizing tumor severity, capturing nuances missed by binary classification [17]. Texture features reflecting tumor heterogeneity and shape irregularity emerged as strong predictors. Diffusion metrics contributed additional microstructural information, reinforcing the importance of multimodal integration [18,19]. The regression approach provided accurate predictions and clinical insights, particularly when supported by explainability techniques like SHAP. Differences in hemispheric tumor location between LGG and HGG further support the biological underpinnings captured by the radiomic features [20].

Radiomic features from T1-weighted and DTI sequences, analyzed through ML regression models, enable accurate, interpretable, and non-invasive glioma grading. The XGBoost regressor demonstrated superior performance and SHAP-based analysis enhanced transparency. This continuous-scale grading strategy offers a refined perspective on tumor biology and supports its use in clinical decision-making. Future studies with larger and multi-site cohorts are encouraged to validate the model’s generalizability and explore its integration into routine radiological workflows. All radiomic processing scripts, ML models, and anonymized datasets used in this study are available from the corresponding author upon reasonable request. Feature extraction was conducted using PyRadiomics v3.0 in Python 3.8, and model development used open-source libraries, including Scikit-learn and XGBoost.

Cortical lesion (CL) detection in magnetic resonance imaging (MRI) is a key biomarker for differential diagnosis and disability assessment in multiple sclerosis (MS) [1,2] (Figure 1). Recent deep learning (DL) methods have improved the standardization and accuracy of CL segmentation [3,4]. Trustworthy AI, particularly with uncertainty quantification (UQ), enhances model reliability and robustness, which are critical for clinical use [5]. While UQ is often used to estimate prediction errors, its clinical interpretation remains underexplored. Our prior studies [6,7] linked high lesion-level uncertainty with interpretable imaging features. This study examines how lesion-scale uncertainty relates to expert perception, including lesion type, diagnostic confidence, and segmentation quality.

The expert perception was assessed with the participation of an experienced neurologist (AC, 6 years of MRI experience). Figure 2 outlines the study workflow. We used our previous 3D nnU-Net model [7,8] trained on a clinical MP2RAGE dataset [9,10] with manual CL annotations by consensus between two experts (a medical doctor and the same neurologist; both had 5 years of MRI experience). The dataset included 163 MS patients (train:val:test = 109:13:41; CLs = 857:69:301). Test set lesion uncertainty was estimated using deep ensembles [11], and lesion structural uncertainty (LSU) [12] was computed (range: 0–1; higher - more uncertain). Lesions were grouped into five equidistant LSU bins; up to 23 per bin were randomly sampled, yielding 100 regions of interest (ROIs; Figure 2, Step 2). Due to the skewed LSU distribution, 68 ROIs had LSU > 0.19 (above 75th percentile); 34 of 100 were false positives. The neurologist reviewed each ROI overlaid on MP2RAGE using their preferred viewer, with detailed written instructions. For each ROI, they reported lesion type, confidence (5-point Likert), reasons for reduced confidence (from 13 predefined categories [7]), and segmentation quality (high/moderate/low) using a multiple-choice table-like form (Figure 2, Step 3). We analyzed categorical expert ratings and continuous variables not visible to the neurologist: LSU and lesion segmentation quality, measured by IoUadj between ROI and ground truth masks [13]. Distributions were visualized via violin plots, and group differences were tested using the Mann–Whitney U-test (n > 10). Inverse confidence scores weighted confidence-reducing factors to highlight dominant sources of uncertainty.

The average self-reported assessment time was 30 sec/ROI. Most ROIs were identified as leukocortical (n=36) or juxtacortical (n=50), with no significant LSU difference between groups (Figure 3-i). Notably, 9 of 34 false positives were reclassified as true lesions. Among the 66 lesions overlapping with the ground truth, 35 were labeled as non-CL (33 juxtacortical MS, 2 vascular MS lesions). Annotator confidence was most often "slightly" (n=42) or "moderately" confident (n=34) (Figures 3-ii, 4-ii). LSU has significantly different distributions across the perceived segmentation quality (Figure 3-iii) with p < 0.01. There was no statistical difference between evaluated segmentation quality (IoUadj) distributions for “high” vs “moderate” perceived quality (p=0.11), but a clear difference between “moderate” vs “low” (p<0.01). By far, the most influential factor lowering confidence was "unclear cortical involvement" (Figure 4-i).

We sampled lesions uniformly across LSU values to explore a possible one-to-one mapping with annotator confidence levels. However, the estimated annotator confidence distribution is also skewed, suggesting that this direct mapping is not viable. On the other hand, most of the sampled ROI had high LSU (above the 75th percentile) and were also deemed difficult by the neurologist, who was mostly “moderately confident” and changed his opinion about 44 of 100 ROIs. "Unclear cortical involvement" emerged as the primary reason for diagnostic uncertainty, consistent with the known challenges of CL detection in 3T MRI [14] and our prior uncertainty-focused work [7]. The observed association between LSU and perceived segmentation quality corroborates previous findings [12] and supports its utility in clinical settings.

This study advances the understanding of how predictive uncertainty aligns with expert confidence in MS cortical lesion segmentation. Our results demonstrate that lesion structural uncertainty from deep ensembles is a promising tool for identifying clinically ambiguous lesions and guiding expert review without ground truth, showing a potential for clinical decision support and targeted review. Future work should expand expert participation to validate these findings and enhance generalizability.

Quantitative MRI provides increasingly relevant biomarkers for the characterization and monitoring of the musculoskeletal system. Such quantitative image analyses generally require delineating regions of interest (ROI) to separate the different muscles or muscle groups. Lately, AI models achieved excellent segmentation quality in a small fraction of the time taken by manual segmentation. Despite their efficiency, these automatic methods often generate partially erroneous segmentations, especially in the presence of important pathological involvements such as fatty replacements. In this case, the only way to correct errors is by manual segmentation, which can be very time-consuming depending on the size of the image and the number of errors. Based on the work of Diaz Pinto et al.[1], we propose a method for semi-automatic 3d muscle segmentation where users can correct errors by providing guidance to the network in the form of clicks or scribbles. On a database of MRI acquisitions or healthy and pathological legs, we trained and tested a modified nnU-net[2] model, including automatically generated clicks within the training loop, and compared its performance to that of the unmodified fully automatic nnU-net model. Initial results show that the proposed model provides a fully automatic segmentation comparable to the non-interactive nnU-net model and, in addition, allows efficient improvements through simple user interactions.

For this study, both training and test sets were composed of 20 MRI acquisitions of legs of consenting volunteers and neuromuscular disease patients, acquired at 3T (PrismaFit, Siemens Healthineers). Images were 3d Dixon acquisitions of 64 slices of 224 x 224 pixels, with TE = 3.95 ms. In both datasets, we selected equal numbers of cases with low and high degrees of fatty replacements. Ground truths were manually delineated for 9 labels : background, tibialis anterior (TA), extensor digitorum (ED), peroneus (PER), deep layer of posterior compartment (DLPC), popliteus (POP), soleus (SOL), lateral gastrocnemius (GAS_LAT), medial gastrocnemius (GAS_MED), by medical experts in 1 out of 4 slices. Using the nnU-Net framework, we modified the training pipeline to introduce simulated clicks using a method similar to that of Sakinis et al.[3]. In each training batch, clicks are progressively and randomly added into dedicated input channels with probabilities depending on the size of the mislabeled regions. Following Diaz-Pinto et al.[1], simulated clicks are added in only 80 % of the batches to maintain performance without click. We also used a modified loss (Figure 1) that forces the network to focus at first on the click-less segmentation and to prioritize click influence in later epochs. We trained our model on a Nvidia RTX A6000 using 5-folds cross validation, with each fold running for 800 epochs each. An example of the incremental correction with manual clicks of a test set image is given in Figure 2. For the testing part, we simulated 20 successive clicks based on the ground truth, computing relevant metrics for each prediction for all labels. The measured metrics were: the Dice similarity coefficient (mean, per label mean, mean of worst label), the maximum error distance (the distance to the border of the innermost pixel of the mislabeling error mask) and the click influence (the number of updated pixels). We also trained and tested an unmodified nnU-Net on the same dataset as benchmark (the “no-click” model). Comparison between models were assessed with paired t-tests.

Figures 3 and 4 show that the 0-click and no-click models have equivalent performance in terms of maximum error distance, Dice of worst label, mean Dice, and per-label Dice. As the number of clicks increase, the mean Dice and maximum error distance become significantly better than for the no-click model. The label frequency plot shows that the larger regions (background, DLPC, SOL) tend to be prioritized for the clicks, while smaller muscles are less often selected. The click influence plot indicates that initial clicks tend to have more influence than later clicks, the latter converging towards about 500 pixels.

These results show that adding corrective guidance to the nnU-net model does not penalize the fully automatic segmentation, while automatically added clicks significantly improve the segmentation quality. We used the nnU-Net framework for practical reasons, but this method can be used with any neural network architecture. Alternative architectures such as transformers could potentially improve segmentation performance. The above results were based on simulated clicks. Using actual user-guidance provided by medical experts is the logical next step to prove the relevance of the proposed method.

After further validation, this approach could be implemented into a simple user-interface to lighten the workload in research and clinical settings, making it possible to generate and correct a full 3D image segmentation in a few minutes.

Dynamic contrast-enhanced (DCE) MRI helps detect and characterize diseases, like cancer and neurodegenerative disorders, by quantifying tissue perfusion. However, conventional methods often yield noisy parameter estimates. Recently, deep learning has emerged as an alternative, offering more accurate and precise parameter estimates. However, even in regions of uncertainty, these deep-learnt parameter-maps are visually appealing , leading clinicians into a false sense of security. Incorporating uncertainty estimates in a deep learning framework can inform clinicians about trustworthiness. Moreover, understanding whether uncertainties stem from noise in the data (aleatoric) or model limitations (epistemic) can inform us whether we need better data or models. For example, in case of out-of-distribution (OOD) effects, epistemic uncertainty may be reduced by offering more varied training examples. Therefore, we propose an uncertainty-aware neural network, hypothesising it can detect aleatoric uncertainties (H1) and that a deep ensemble can capture epistemic uncertainties (H2). Finally, we demonstrate how this model’s uncertainties can be visualized in vivo.

We used the extended Tofts model to simulate 100,000 pharmacokinetic concentration-time curves with varying noise levels. This synthetic dataset was split into training, validation, and test subsets. To test H1, we extended a previously developed network (DCE-NET) by incorporating mean-variance estimation (MVE), enabling the model to output both pharmacokinetic parameters and associated uncertainty. We evaluated whether predicted uncertainties correlated with actual errors using quartile-based stratification and statistical comparisons of error distributions. To test H2, we trained an ensemble of ten independently initialized MVE-DCE-NET models. Each network’s output was treated as a sample from a posterior distribution, allowing us to compute both predictive mean and epistemic uncertainty. To simulate OOD effects, we intentionally excluded certain parameter ranges from the training and validation data while keeping them in the test set. This allowed us to observe whether the ensemble detected unfamiliar parameter regimes through increased uncertainty. Finally, we developed a visualization method for uncertainty-aware parameter maps. These maps display predicted parameter values and visually encode confidence by superimposing controlled noise that scales with predicted uncertainty. As a proof of concept, we applied our method to in vivo DCE-MRI data from two healthy volunteers.

The uncertainty-aware networks showed more accurate pharmacokinetic parameter predictions than the conventional network (Figure 1A). Larger aleatoric uncertainties predicted by MVE-DCE-NET correlated with a larger spread in parameter errors (Figure 1B). Paired Levene’s test indicated significant differences in error variances across uncertainty quartiles, confirming H1. Additionally, a positive correlation between median predicted uncertainty and error further supported this hypothesis (Figure 2). Evaluation of epistemic uncertainties revealed that the ensemble effectively identified unseen parameters in the test dataset (Figure 3), with lowest uncertainties for well-represented values and highest for unrepresented ones. Uncertainty was also high for low ke and ve. As curves with little tissue uptake can be explained both with a low ke (and arbitrary ve) or low ve (and arbitrary ke), there is a large uncertainty on ke and ve in those situations. With no noise, in vivo parameter maps look smooth and trustworthy (Figure 4). However, as the noise scaling parameter increased, the uncertainty-informed maps displayed more noise in uncertain regions, making it clear to the viewer which regions are less trustworthy.

This study presents MVE-DCE-NET, an uncertainty-aware neural network designed to predict pharmacokinetic parameters along with their aleatoric and epistemic uncertainties. Our findings demonstrate that MVE-DCE-NET effectively highlights high aleatoric uncertainties when estimation errors are large and shows low uncertainty when errors are low, making results interpretable. Additionally, deep ensembles identified underrepresented data with high uncertainties, which could support create balanced datasets with equal representations of healthy and diseased tissues. If certain tissues exhibit parameters not present in the training data, our approach can flag these tissues, encouraging further examination and reducing reliance on uncertain predictions.

This work introduces MVE-DCE-NET, an uncertainty-aware model that enhanced DCE-MRI parameter reliability, distinguishing between inherent data noise and model limitations. This approach can support clinicians in making informed decisions by highlightening prediction confidence and potential outliers.

Magnetic resonance-guided radiotherapy (MRgRT) enables superior soft-tissue visualisation and daily adaptive treatment planning. Online adaptation can extend treatment sessions by 20–30 minutes, with high-resolution imaging potentially further contributing to this delay [1, 2]. Apparent diffusion coefficient (ADC) maps derived from diffusion-weighted imaging (DWI) show promise as a reliable biomarker for dose escalation based on tumour response [3]. Strategies to accelerate acquisition while preserving image quality are key for improving workflow efficiency for MRgRT where imaging enables online treatment adaptation, such as identifying targets for dose escalation. Although current workflows may accommodate longer scans, future optimizations in radiotherapy delivery will require that imaging speed also keep pace. This study aims to accelerate diffusion imaging for MRgRT using deep learning to generate high-quality DWI and assess the bias and repeatability of the resulting quantitative ADC maps.

This study analysed DWI from 11 prostate cancer patients enrolled in the HERMES clinical trial (REC 20/LO/1162) [4], treated on a 1.5T Unity MR-Linac (Elekta AB, Stockholm, Sweden). The cohort included patients from both the two-fraction (2#) (P1 – P6) and five-fraction (5#) (P7 – P11) arms, with a median age of 72 years (range: 53–77). All patients received neoadjuvant and concurrent androgen deprivation therapy. DW images were acquired using single-shot echo planar imaging with b-values(averages) of 0(6), 30(6), 150(6), and 500(14) s/mm² based on an MR-Linac consortium recommended consensus protocol [5]. A deep learning model [6] based on a U-Net architecture was trained to produce high-quality DWI from noisy input data [Fig. 1]. Training was performed over 70 epochs using the Adam optimizer (learning rate 1e−4), with mean absolute error (MAE) as the loss function and a batch size of 20. Each training scan consisted of 15 transverse slices, with 30 training samples per slice assembled from different diffusion averages, resulting in 450 training samples per scan. Each sample was represented as a [224 × 224 × 12] volume, where the 12 input channels comprised the current slice, preceding and following slices, and images from the four b-values. For the models, the input at each b-value was a directionally averaged trace image. The specific acquisition or means of any two or three acquisitions was/were randomly selected per sample. In total, 25 scans (11,250 samples) were used for training and 5 scans (2,250 samples) for validation. A five-fold cross-validation was used for the best model. For Fold 1, patient P1 was used for testing, P6 and P10 for validation, and the remaining patients (P2–P5, P7–P9, and P11) for training. For the subsequent training folds, the other 2# patient data were rotated to be test data. Three model configurations were evaluated, differing in diffusion input averaging: all-direction input using single, two and three averages respectively: 1. AD_avg1 2. AD_avg2 3. AD_avg3 Quantitative evaluation included voxel-wise fit of ADC of whole prostate (WP) and comparison between model outputs and ground truth (GT). The ADC repeatability coefficients (RC) of WP and gross-tumour volume (GTV) were calculated.

Fig. 2 shows a representative transverse slice for visual comparison of GT, inputs and the predicted images. MAE and RMSE between predicted and GT ADC are reported in Table 1. AD_avg3 had the lowest error, followed by AD_avg2. Box plots of WP ADC distributions showed good agreement between model predictions and GT, with the models showing a narrower distribution [Fig. 3]. AD_avg2 had a negative bias for the tested patient. RC Calculation: Absolute ADC RCs in 10-6 mm2/s (relative RC) for the full acquisition were 483 (28.0%), and 234 (12.8%) for GTV and whole prostate respectively. In the predicted ADC maps of AD_avg1 model, these were improved to 181.83 (10.4%) and 180.48 (9.5%).

We demonstrate that deep learning can generate high-quality DWI from subsampled data on the MR-Linac. Specifically, model trained with only one average from all-direction data (AD_avg1) produced results visually comparable to full averaging, reducing acquisition time to 36 seconds from the current 4:12 minutes. This model achieved superior ADC repeatability compared to the full acquisition ground truth for this dataset. There is a clear trade-off between denoising performance and acquisition time. While AD_avg3 achieves the lowest error, AD_avg1 offers substantial scan time reduction with only a marginal increase in error, suggesting it may be preferable in a time-constrained clinical setting.

Deep-learning accelerated DWI makes feasible shorter scan times, improved patient comfort, and better workflow compatibility for diffusion imaging in MRgRT without compromising image quality.

The use of deep learning has increased significantly in recent years in the field of neuroscience. Various methods are being developed as aid-to-diagnosis tools from imaging data with successful results. Our objective is to study how different parametric maps, all calculated from T1-weighted (T1w) magnetic resonance imaging (MRI), influence the learning process of a 3D convolutional neural network (3D CNN) to classify healthy controls (HC) from patients suffering from multiple system atrophy (MSA), a rare neurodegenerative disease [1].

The dataset comprised T1w MRI acquisitions gathered from three MSA reference centres in France [1-5], including 126 HC and 92 MSA patients. Images were processed using SPM12 on MATLAB R2019b with the following steps: (i) spatial normalization of T1w volumes in the Montreal National Institute (MNI) space (ii) resampling to a 2×2×2 mm3 resolution; (iii) skull stripping; (iv) segmentation of brain substances from the normalized T1w, obtaining grey matter density (GD), white matter density (WD) and cerebrospinal fluid (CSF) maps. We considered a 3D CNN architecture [4] to perform a binary classification task between HC and MSA patients, using a single image type (monomodal) and different combinations from the maps (bimodal, trimodal, and quadrimodal). The CNN was trained for 30 epochs with a ten-time repeated five-folds cross-validation with two different dataset splits. The database was divided into 80%/20% for training/testing our model. We considered accuracy, specificity and sensibility on the hold-out sets to evaluate CNN performance according to the different inputs.

In the monomodal study, we obtained the best mean accuracy with GD and WD maps (0.92 ± 0.06 and 0.94 ± 0.06 respectively), followed by T1 and CSF (0.87 ± 0.05 and 0.88 ± 0.04). However, the parametric maps showed a greater difference between sensitivity (0.50-1.00) and specificity (0.89-1.00), compared to the model based on T1 images (T1-CNN). The multimodal study showed an overall improvement in performance, with a mean accuracy of 0.90 ± 0.06 for the T1-GD-CNN and particularly for the T1-GD-WD-CNN, with a mean accuracy of 0.93 ± 0.06. The T1-GD-WD-CSF-CNN yielded results comparable to those of the previously mentioned models with a 0.92 ± 0.06 accuracy.

Firstly, the monomodal study enabled us to demonstrate the good classification capabilities of our 3D CNN to distinguish MSA patients from HC. Second, the multimodal study allowed us to observe a general improvement in performance depending on the type of map used with the T1 images. In our bimodal study, we found that adding as input a single parametric map improves the classification capabilities of the 3D CNN, but also can inform about the interpretability of the model from the monomodal performance. For example, considering the T1-GD-CNN, we can suppose that performances will benefit from the better classification found with GD maps alone. In the trimodal approach, we observe an increase in classification capabilities but interpreting these models becomes more challenging as we cannot know to which extent each map contributes to the multimodal performance. Finally, our quadrimodal study yielded classification capabilities comparable to the trimodal study, while increasing computational complexity. It is also important to note that there is a trade-off between the classification capabilities of our model and the computational cost. Indeed, the addition of a modality greatly increases the computational time of the algorithm. Although our sample size is limited by the rarity of the disease, these results are encouraging and show that multimodality can play a role in enhancing diagnostic accuracy. Lastly, to our knowledge, no study has yet exploited the unique contribution of parametric maps derived from the same MRI sequence targeting different aspects of brain structure compared to the sequence itself. This approach could pave the way for a better understanding of the “black box” nature of 3D CNNs thanks to the analysis of input data. However, the use of visualization techniques and the study of misclassified patients could help us to shed light on the impact of parametric maps on the 3D CNN performance.

Using different T1-derived maps as input to a 3D CNN increased classification performance to distinguish HC from MSA patients. However, the monomodal study based on the parametric maps calculated from the T1-weighted sequence showed the contribution of each cerebral substance by providing more precise information. To validate our approach, we plan to apply it to other pathologies, such as Alzheimer's disease, and to extend it to the differential diagnosis of parkinsonian syndromes.

Magnetic resonance imaging (MRI) of the human spinal cord (SC) has seen significant technical advances, enabling the visualization of white matter (WM) and gray matter (GM) across various sequences such as MGE-T2starw, PSIR [1], TSE-T1w, MTR, rAMIRA [2], MP2RAGE [3], and SWI [4], at different field strengths (1.5T, 3T, 7T). GM segmentation is a key step for extracting biomarkers such as cross-sectional area (CSA) [5,6], which are essential for monitoring neurodegenerative and traumatic pathologies including MS [7,8], ALS [2,9,10], SCI [11], SMA [12], PPS [13], and DCM [14]. It also facilitates enhanced image registration, inter-subject alignment within the lumbosacral SC [15], tensor-based morphometry [16], ROI-based qMRI analyses [17], new templates construction [18] and fMRI processing [19–21]. Manual segmentation is time-consuming and subject to inter- and intra-rater variability [15,22,23], prompting the development of automated approaches [5,16,22,24–26]. Among these, the deep learning-based model sct_deepseg_gm [27], integrated into the Spinal Cord Toolbox (SCT) [28], has demonstrated strong performance particularly on MGE-T2starw images and similar contrasts and was the winner of the SC GM Segmentation Challenge [22]. However, the growing diversity of MRI contrasts, acquisition protocols, and use cases, including pediatric and thoraco-lumbar imaging, presents ongoing challenges for model generalization. This study aims to develop an automatic GM segmentation method that is robust across contrasts, field strengths, SC regions, and pathologies.

Data This study relies on multicentric data from 16 sites (Figure 1a), including 3 magnetic field strengths and 12 MRI contrasts grouped into 9 categories (Figure 1b). The dataset includes pediatrics controls, adult healthy controls (HC) as well as patients with ALS, MS, SMA, PPS, DCM, SCI, and stroke, for a total of 1,367 volumes (Figure 1c). Preprocessing All images were reoriented to RPI orientation and resampled in the axial plane to 0.3x0.3 mm2 using SCT v6.5 [28]. GM Ground Truth Manual GM segmentations, provided as binary masks, were available from 10 sites. For the remaining 6 sites, GM segmentation was first obtained using a preliminary version of our method (r20250204), followed by manual correction to ensure accuracy. Automated Segmentation Benchmark: Contrast-specific models were used as benchmarks (Table 1). For contrasts without existing segmentation methods (PSIR, 3T-TSE-T1w, 7T-MP2RAGE-T1map, and thoraco-lumbar images), single-contrast models were trained using the nnU-Net v2 framework [29]. Our approach: The nnUNetv2 framework was used to train on 1000 epochs in 5 folds with cross-validation on a 2D model. The training dataset consisted of 1,141 volumes (15,535 axial slices). The testing dataset comprised 20% of the images from sites with GTs (Figure 1a). Evaluation of Segmentation Segmentation performance was assessed using the Dice Similarity Coefficient [30] and the Hausdorff Distance (HD) in mm [31], calculated for each 2D axial slice. Statistics Wilcoxon test for paired samples, comparing benchmark and our approach.

Figure 2 summarizes the Dice scores and HD, showing that our approach significantly outperforms the benchmark methods in multiple sclerosis (MS) cases (p < 0.05). This demonstrates the model’s superior performance in segmenting gray matter in the presence of MS lesions (Figure 3). A Dice score above 0.85 (Figure 2d) indicates high segmentation fidelity, comparable to that of benchmark #1 [27] reported in the SC GM Segmentation Challenge. However, benchmark #1 shows reduced performance in the lower cervical and thoracic regions (Figure 2a,b). For the 7T datasets, both the benchmarks (#3, #6) and our method performs well.

While our method and the benchmarks perform adequately overall, segmentation in patients with SMA and DCM is less accurate (Figure 2a), likely due to SC compression that alters GM shape and contrast. This highlights the need for more training data from varied pathological cases. In the thoraco-lumbar region (Figure 2b), our model maintains good performance, although segmentation accuracy decreases at the L2 level due to the small size of the spinal cord at this point (~10 pixels; Figure 3). Dice scores are impacted by the limited number of GM voxels in this region across all methods. Our model achieves an average HD below 0.6 mm across most contrasts, indicating that the predicted GM contours typically deviate by only one pixel from the ground truth. This robustness holds across diverse GM morphologies and pathological conditions (Figure 3)

We present a robust approach for automatic GM segmentation, demonstrating consistent performance across multiple MRI contrasts, SC regions, and pathological conditions. Our approach is publicly available in the SCT v7.0. To further enhance generalizability, especially in challenging pathological cases, future work should incorporate more diverse and pathology-rich datasets during training

AI-based MR reconstructions can accelerate scanning and/or improve image quality. Guidelines for deploying AI tools in radiology[1–4] recommend post-deployment monitoring, as performance may vary due to scanner software updates[5], yet lack specific methods for monitoring that follow established quality control (QC) frameworks: defining performance metrics, baseline values, monitoring frequency, and feedback mechanisms[6–8]. Radiologist assessments are resource-intensive; while phantom-based methods have not been developed for AI-based reconstructions, making both unsuitable for QC. Image quality metrics (IQMs), used to evaluate performance of AI-based images and reconstructions[9–14], offer a more efficient alternative for monitoring. Our institution introduced an AI-based reconstruction for rectal MRI, evaluated using radiologist scoring[15]. We developed a QC programme using IQMs proven to be sensitive to known perturbations in the AI-based reconstruction[16]. The aim of this study was to evaluate 24 months of QC data, during which the MR systems underwent software upgrades with known modifications to the reconstruction pipeline.

90 patients (47M/43F, 33-89y, median 64y) undergoing routine anorectal MRI (Apr 2023-Apr 2025) were scanned on a 1.5T or 3T system (1.5T-A: n=19, 1.5T-B: n=18, 3T-A: n=36 and 3T-B: n=17, MAGNETOM Sola/Vida, Siemens Healthineers, Forchheim, Germany) using a 30-ch body array and 32-ch spine array. The study was approved by a research ethics committee; patients gave verbal consent for additional imaging as part of routine clinical MRI examinations. Axial T2-weighted small field-of-view turbo spin-echo sequences were acquired with and without AI-based reconstruction[15, 16] at matched positions. Raw data were also saved. Patients were numbered chronologically, the first 50 patients formed the reference dataset; the rest formed the constancy dataset. During the study, the 1.5T-A, 3T-A and 3T-B systems underwent software upgrades. A radiologist qualitatively reviewed 3 prospective post-upgrade datasets per upgraded system. Post-upgrade reconstructions and difference images were generated for 3 datasets per 3T system (Fig.1). Paired IQMs – mean squared error (MSE), peak signal-to-noise ratio (pSNR), structural similarity index (SSIM), and visual information fidelity (VIF)[17–19] – and unpaired IQMs – image entropy, and Tenengrad[17, 20] – were evaluated in a 214×214-pixel central region-of-interest for all slices. IQM values for all datasets were evaluated using Shewhart control charts, comparing each exam's mean IQM against the reference mean and standard deviation (SD). Control chart test rules (Fig.2) were applied[21]. Radiologist scoring was performed for the reference dataset[15]. IQMs were evaluated using MATLAB (R2024b, Natick, MA).

Additional QC data acquisition took an average of 6 minutes per exam. Difference images (Fig.1B) showed changes in accelerated acquisitions with a mean pixel intensity of 0.4 ± 0.6 whereas standard acquisitions showed minimal change, with a mean pixel intensity of 0.0007 ± 0.04. Mean IQM values for patients 56 (SSIM, VIF), 60 (SSIM) and 85 (MSE, pSNR, SSIM, VIF) exceeded 2 SDs, triggering control chart rule 1 (Figs.2&3). Rules 2-4 were not triggered. Mean IQM values for post-upgrade retrospective reconstructions did not trigger any control chart rules. Only rule 1 was triggered for patients 56 (SSIM, 1.5T-A); 60 (MSE, pSNR, SSIM 3T-A) and 85 (MSE, pSNR, SSIM, VIF, 1.5T-A) when evaluating the IQMs per scanner (Fig.4).

This study evaluated IQMs for QC of AI-based MR reconstructions over 2 years during which there were system software upgrades. Longitudinal IQM trends indicated stable performance since clinical deployment as no results triggered additional inspection of data or re-evaluation of the imaging. Examinations with outlier IQM values were traced to gross motion. Motion was generally minimised through patient setup and use of antispasmodics. No IQM values indicated changes outside of expected variation for examinations conducted post software upgrade, suggesting image quality is consistent with reference data for which radiologist scoring indicated good image quality. Differences images showed that software upgrades altered the reconstruction pipeline - specifically, a change in the phase correction method for the AI-based reconstruction. Despite this, neither IQMs nor radiologist review indicated changes in image quality. The IQMs provide a resource-efficient QC method, with automated, quantitative analysis with minimal added acquisition time and no extra radiologist burden.

IQMs offer a resource-efficient alternative to routine radiological evaluation for ongoing monitoring of AI-based MR reconstructions. No clinically significant performance changes were detected by IQMs or radiologists after software upgrades. Future work will include radiologist scoring of post-upgrade data and expansion of this QC framework to other MR sequences.