Parkinson's disease (PD) is a neurodegenerative disorder that affects both motor and cognitive functions, particularly working memory (WM). Machine learning offers an advantage for decoding complex brain activity patterns, but its application to task-based functional magnetic resonance imaging (task-based fMRI) has been limited. This study aimed to develop an explainable machine learning model to classify WM performance levels in PD based on task-based fMRI data. We enrolled 45 patients with PD and 15 healthy controls (HCs), all of whom performed an n-back WM task in an MRI scanner. Patients were stratified into three subgroups based on their 3-back task performance: better, intermediate, and worse WM. A three-dimensional convolutional neural network (3D-CNN) model, pre-trained with a 3D convolutional autoencoder, was developed to perform binary classifications between group pairs. The model achieved an accuracy of 93.3% in discriminating task-based fMRI images of PD patients with worse WM from HCs, surpassing the mean accuracy of three expert radiologists (70.0%). Saliency maps identified brain regions influencing the model's decisions, including the dorsolateral prefrontal cortex and superior/inferior parietal lobules. These regions were consistent with both the areas with intergroup differences in the task-based fMRI data and the anatomical areas that are crucial for better WM performance. We developed an explainable deep learning model that is capable of classifying WM performance levels in PD using task-based fMRI. This approach may enhance the objective and interpretable assessment of brain function in clinical neuroimaging practice.

This study aimed to develop and evaluate a Transformer-CNN framework for automated segmentation of multiple sclerosis (MS) lesions on FLAIR MRI. The model was benchmarked against U-Net and DeepLabV3 and assessed for both segmentation accuracy and across-center performance under internal 5-fold cross-validation to ensure robustness across diverse clinical datasets. A dataset of 1,800 3D FLAIR MRI scans from five clinical centers was split using 5-fold cross-validation. Preprocessing included isotropic resampling, intensity normalization, and bias field correction. The Transformer-CNN combined CNN-based local feature extraction with Transformer-based global context modeling. Data augmentation strategies, including geometric transformations and noise injection, enhanced generalization. Performance was evaluated using Dice score, IoU, HD95, and pixel accuracy, along with internal cross-validation-based metrics such as Generalized Dice Similarity Coefficient (GDSC), Domain-wise IoU (DwIoU), Cross-Fold Dice Deviation (CFDD), and Volume Agreement (Intraclass Correlation Coefficient, ICC). Statistical significance was tested using Kruskal-Wallis and Dunn's post-hoc analyses to compare models. The Transformer-CNN achieved the best overall performance, with a Dice score of 92.3%, IoU of 91.4%, HD95 of 2.25 mm, and pixel accuracy of 95.6%. It also excelled in internal cross-validation-based across-center metrics, achieving the highest GDSC (91.3%) and DwIoU (89.2%), the lowest CFDD (1.05%), and the highest ICC (96.5%). DeepLabV3 and U-Net scored 85.1% and 83.0% in Dice, with HD95 values of 4.15 mm and 4.30 mm, respectively. The worst performance was observed in U-Net, which exhibited high variability across datasets and struggled with small lesion detection. The Transformer-CNN outperformed U-Net and DeepLabV3 in segmentation accuracy and across-center performance under internal 5-fold cross-validation. Its robustness, minimal variability, and ability to generalize across diverse datasets establish it as a practical and reliable tool for clinical MS lesion segmentation and monitoring.

Transformer-based deep learning models have demonstrated exceptional performance in medical imaging by leveraging attention mechanisms for feature representation and interpretability. However, these models are prone to learning spurious correlations, leading to biases and limited generalization. While human-AI attention alignment can mitigate these issues, it often depends on costly manual supervision. In this work, we propose a Hybrid Explanation-Guided Learning (H-EGL) framework that combines self-supervised and human-guided constraints to enhance attention alignment and improve generalization. The self-supervised component of H-EGL leverages class-distinctive attention without relying on restrictive priors, promoting robustness and flexibility. We validate our approach on chest X-ray classification using the Vision Transformer (ViT), where H-EGL outperforms two state-of-the-art Explanation-Guided Learning (EGL) methods, demonstrating superior classification accuracy and generalization capability. Additionally, it produces attention maps that are better aligned with human expertise.

This paper introduces a novel framework for image quality transfer based on conditional flow matching (CFM). Unlike conventional generative models that rely on iterative sampling or adversarial objectives, CFM learns a continuous flow between a noise distribution and target data distributions through the direct regression of an optimal velocity field. We evaluate this approach in the context of low-field magnetic resonance imaging (LF-MRI), a rapidly emerging modality that offers affordable and portable scanning but suffers from inherently low signal-to-noise ratio and reduced diagnostic quality. Our framework is designed to reconstruct high-field-like MR images from their corresponding low-field inputs, thereby bridging the quality gap without requiring expensive infrastructure. Experiments demonstrate that CFM not only achieves state-of-the-art performance, but also generalizes robustly to both in-distribution and out-of-distribution data. Importantly, it does so while utilizing significantly fewer parameters than competing deep learning methods. These results underline the potential of CFM as a powerful and scalable tool for MRI reconstruction, particularly in resource-limited clinical environments.

One-shot medical image segmentation faces fundamental challenges in prototype representation due to limited annotated data and significant anatomical variability across patients. Traditional prototype-based methods rely on deterministic averaging of support features, creating brittle representations that fail to capture intra-class diversity essential for robust generalization. This work introduces Diffusion Prototype Learning (DPL), a novel framework that reformulates prototype construction through diffusion-based feature space exploration. DPL models one-shot prototypes as learnable probability distributions, enabling controlled generation of diverse yet semantically coherent prototype variants from minimal labeled data. The framework operates through three core innovations: (1) a diffusion-based prototype enhancement module that transforms single support prototypes into diverse variant sets via forward-reverse diffusion processes, (2) a spatial-aware conditioning mechanism that leverages geometric properties derived from prototype feature statistics, and (3) a conservative fusion strategy that preserves prototype fidelity while maximizing representational diversity. DPL ensures training-inference consistency by using the same diffusion enhancement and fusion pipeline in both phases. This process generates enhanced prototypes that serve as the final representations for similarity calculations, while the diffusion process itself acts as a regularizer. Extensive experiments on abdominal MRI and CT datasets demonstrate significant improvements respectively, establishing new state-of-the-art performance in one-shot medical image segmentation.

We present CuMPerLay, a novel differentiable vectorization layer that enables the integration of Cubical Multiparameter Persistence (CMP) into deep learning pipelines. While CMP presents a natural and powerful way to topologically work with images, its use is hindered by the complexity of multifiltration structures as well as the vectorization of CMP. In face of these challenges, we introduce a new algorithm for vectorizing MP homologies of cubical complexes. Our CuMPerLay decomposes the CMP into a combination of individual, learnable single-parameter persistence, where the bifiltration functions are jointly learned. Thanks to the differentiability, its robust topological feature vectors can be seamlessly used within state-of-the-art architectures such as Swin Transformers. We establish theoretical guarantees for the stability of our vectorization under generalized Wasserstein metrics. Our experiments on benchmark medical imaging and computer vision datasets show the benefit CuMPerLay on classification and segmentation performance, particularly in limited-data scenarios. Overall, CuMPerLay offers a promising direction for integrating global structural information into deep networks for structured image analysis.

Foundation models open up new possibilities for the use of AI in healthcare. However, even when pre-trained on health data, they still need to be fine-tuned for specific downstream tasks. Furthermore, although foundation models reduce the amount of training data required to achieve good performance, obtaining sufficient data is still a challenge. This is due, in part, to restrictions on sharing and aggregating data from different sources to protect patients' privacy. One possible solution to this is to fine-tune foundation models via federated learning across multiple participating clients (i.e., hospitals, clinics, etc.). In this work, we propose a new personalized federated fine-tuning method that learns orthogonal LoRA adapters to disentangle general and client-specific knowledge, enabling each client to fully exploit both their own data and the data of others. Our preliminary results on real-world federated medical imaging tasks demonstrate that our approach is competitive against current federated fine-tuning methods.

Deep learning relies heavily on data augmentation to mitigate limited data, especially in medical imaging. Recent multimodal learning integrates text and images for segmentation, known as referring or text-guided image segmentation. However, common augmentations like rotation and flipping disrupt spatial alignment between image and text, weakening performance. To address this, we propose an early fusion framework that combines text and visual features before augmentation, preserving spatial consistency. We also design a lightweight generator that projects text embeddings into visual space, bridging semantic gaps. Visualization of generated pseudo-images shows accurate region localization. Our method is evaluated on three medical imaging tasks and four segmentation frameworks, achieving state-of-the-art results. Code is publicly available on GitHub: https://github.com/11yxk/MedSeg_EarlyFusion.

Although deep neural networks have provided impressive gains in performance, these improvements often come at the cost of increased computational complexity and expense. In many cases, such as 3D volume or video classification tasks, not all slices or frames are necessary due to inherent redundancies. To address this issue, we propose a novel learnable subsampling framework that can be integrated into any neural network architecture. Subsampling, being a nondifferentiable operation, poses significant challenges for direct adaptation into deep learning models. While some works, have proposed solutions using the Gumbel-max trick to overcome the problem of non-differentiability, they fall short in a crucial aspect: they are only task-adaptive and not inputadaptive. Once the sampling mechanism is learned, it remains static and does not adjust to different inputs, making it unsuitable for real-world applications. To this end, we propose an attention-guided sampling module that adapts to inputs even during inference. This dynamic adaptation results in performance gains and reduces complexity in deep neural network models. We demonstrate the effectiveness of our method on 3D medical imaging datasets from MedMNIST3D as well as two ultrasound video datasets for classification tasks, one of them being a challenging in-house dataset collected under real-world clinical conditions.

Following atrial fibrillation ablation, it is challenging to distinguish patients who will remain arrhythmia-free from those at risk for recurrence. New explainable machine learning (xML) techniques allow for systematic assessment of arrhythmia recurrence risk following catheter ablation. We aim to develop an xML algorithm that predicts recurrence and reveals key risk factors to facilitate better follow-up strategy after an ablation procedure. We reconstructed pre-and post-ablation models of the left atrium (LA) from late gadolinium enhanced magnetic resonance (LGE-MRI) for 67 patients. Patient-specific features (LGE-based measurements of pre/post-ablation arrhythmogenic substrate, LA geometry metrics, computational simulation results, and clinical risk factors) trained a random forest classifier to predict recurrent arrhythmia. We calculated each risk factor's marginal contribution to model decision making via SHapley Additive exPlanations (SHAP). The classifier accurately predicts post-ablation arrhythmia recurrence (mean receiver operating characteristic [ROC] area under the curve [AUC]: 0.80 ± 0.04; mean precision-recall [PR] AUC: 0.82 ± 0.08). SHAP analysis reveals that of 89 features tested, the key population risk factors for recurrence are: large left atrium, low LGE-quantified post-ablation scar in the atrial floor region, and previous attempts at direct current cardioversion. We also examine patient-specific recurrence predictions, since xML allows us to understand why a particular individual can have large prediction weights for some categories without tipping the balance towards an incorrect prediction. Finally, we validate our model in a completely new, 15-patient retrospective holdout cohort (80% correct). Our SHAP-based explainable machine learning approach is a proof-of-concept clinical tool to explain arrhythmia recurrence risk in patients who underwent ablation by combining patient-specific clinical profiles and LGE-derived data.