Wilson Silva

Valentina Corbetta, Regina Beets-Tan, Wilson Silva

Multi-centre colonoscopy images from various medical centres exhibit distinct complicating factors and overlays that impact the image content, contingent on the specific acquisition centre. Existing Deep Segmentation networks struggle to achieve adequate generalizability in such data sets, and the currently available data augmentation methods do not effectively address these sources of data variability. As a solution, we introduce an innovative data augmentation approach centred on interpretability saliency maps, aimed at enhancing the generalizability of Deep Learning models within the realm of multi-centre colonoscopy image segmentation. The proposed augmentation technique demonstrates increased robustness across different segmentation models and domains. Thorough testing on a publicly available multi-centre dataset for polyp detection demonstrates the effectiveness and versatility of our approach, which is observed both in quantitative and qualitative results. The code is publicly available at: https://github.com/nki-radiology/interpretability_augmentation

Pedro C. Neto, Tiago Gonçalves, João Ribeiro Pinto et al.

As two sides of the same coin, causality and explainable artificial intelligence (xAI) were initially proposed and developed with different goals. However, the latter can only be complete when seen through the lens of the causality framework. As such, we propose a novel causality-inspired framework for xAI that creates an environment for the development of xAI approaches. To show its applicability, biometrics was used as case study. For this, we have analysed 81 research papers on a myriad of biometric modalities and different tasks. We have categorised each of these methods according to our novel xAI Ladder and discussed the future directions of the field.

Wilson Silva, Maria Carvalho, Carlos Mavioso et al.

Treatments for breast cancer have continued to evolve and improve in recent years, resulting in a substantial increase in survival rates, with approximately 80\% of patients having a 10-year survival period. Given the serious impact that breast cancer treatments can have on a patient's body image, consequently affecting her self-confidence and sexual and intimate relationships, it is paramount to ensure that women receive the treatment that optimizes both survival and aesthetic outcomes. Currently, there is no gold standard for evaluating the aesthetic outcome of breast cancer treatment. In addition, there is no standard way to show patients the potential outcome of surgery. The presentation of similar cases from the past would be extremely important to manage women's expectations of the possible outcome. In this work, we propose a deep neural network to perform the aesthetic evaluation. As a proof-of-concept, we focus on a binary aesthetic evaluation. Besides its use for classification, this deep neural network can also be used to find the most similar past cases by searching for nearest neighbours in the highly semantic space before classification. We performed the experiments on a dataset consisting of 143 photos of women after conservative treatment for breast cancer. The results for accuracy and balanced accuracy showed the superior performance of our proposed model compared to the state of the art in aesthetic evaluation of breast cancer treatments. In addition, the model showed a good ability to retrieve similar previous cases, with the retrieved cases having the same or adjacent class (in the 4-class setting) and having similar types of asymmetry. Finally, a qualitative interpretability assessment was also performed to analyse the robustness and trustworthiness of the model.

Philip Schutte, Valentina Corbetta, Regina Beets-Tan et al.

Federated Learning (FL) in Deep Learning (DL)-automated medical image segmentation helps preserving privacy by enabling collaborative model training without sharing patient data. However, FL faces challenges with data heterogeneity among institutions, leading to suboptimal global models. Integrating Disentangled Representation Learning (DRL) in FL can enhance robustness by separating data into distinct representations. Existing DRL methods assume heterogeneity lies solely in style features, overlooking content-based variability like lesion size and shape. We propose FedGS, a novel FL aggregation method, to improve segmentation performance on small, under-represented targets while maintaining overall efficacy. FedGS demonstrates superior performance over FedAvg, particularly for small lesions, across PolypGen and LiTS datasets. The code and pre-trained checkpoints are available at the following link: https://github.com/Trustworthy-AI-UU-NKI/Federated-Learning-Disentanglement

Aniek Eijpe, Valentina Corbetta, Kalina Chupetlovska et al.

Cross-modal medical image segmentation presents a significant challenge, as different imaging modalities produce images with varying resolutions, contrasts, and appearances of anatomical structures. We introduce compositionality as an inductive bias in a cross-modal segmentation network to improve segmentation performance and interpretability while reducing complexity. The proposed network is an end-to-end cross-modal segmentation framework that enforces compositionality on the learned representations using learnable von Mises-Fisher kernels. These kernels facilitate content-style disentanglement in the learned representations, resulting in compositional content representations that are inherently interpretable and effectively disentangle different anatomical structures. The experimental results demonstrate enhanced segmentation performance and reduced computational costs on multiple medical datasets. Additionally, we demonstrate the interpretability of the learned compositional features. Code and checkpoints will be publicly available at: https://github.com/Trustworthy-AI-UU-NKI/Cross-Modal-Segmentation.

Miriam Cobo, Amaia Pérez del Barrio, Pablo Menéndez Fernández-Miranda et al.

Prognosis after intracranial hemorrhage (ICH) is influenced by a complex interplay between imaging and tabular data. Rapid and reliable prognosis are crucial for effective patient stratification and informed treatment decision-making. In this study, we aim to enhance image-based prognosis by learning a robust feature representation shared between prognosis and the clinical and demographic variables most highly correlated with it. Our approach mimics clinical decision-making by reinforcing the model to learn valuable prognostic data embedded in the image. We propose a 3D multi-task image model to predict prognosis, Glasgow Coma Scale and age, improving accuracy and interpretability. Our method outperforms current state-of-the-art baseline image models, and demonstrates superior performance in ICH prognosis compared to four board-certified neuroradiologists using only CT scans as input. We further validate our model with interpretability saliency maps. Code is available at https://github.com/MiriamCobo/MultitaskLearning_ICH_Prognosis.git.

Aniek Eijpe, Soufyan Lakbir, Melis Erdal Cesur et al.

While multimodal survival prediction models are increasingly more accurate, their complexity often reduces interpretability, limiting insight into how different data sources influence predictions. To address this, we introduce DIMAFx, an explainable multimodal framework for cancer survival prediction that produces disentangled, interpretable modality-specific and modality-shared representations from histopathology whole-slide images and transcriptomics data. Across multiple cancer cohorts, DIMAFx achieves state-of-the-art performance and improved representation disentanglement. Leveraging its interpretable design and SHapley Additive exPlanations, DIMAFx systematically reveals key multimodal interactions and the biological information encoded in the disentangled representations. In breast cancer survival prediction, the most predictive features contain modality-shared information, including one capturing solid tumor morphology contextualized primarily by late estrogen response, where higher-grade morphology aligned with pathway upregulation and increased risk, consistent with known breast cancer biology. Key modality-specific features capture microenvironmental signals from interacting adipose and stromal morphologies. These results show that multimodal models can overcome the traditional trade-off between performance and explainability, supporting their application in precision medicine.

Laura Latorre, Liliana Petrychenko, Regina Beets-Tan et al.

We explore deep generative models to generate case-based explanations in a medical federated learning setting. Explaining AI model decisions through case-based interpretability is paramount to increasing trust and allowing widespread adoption of AI in clinical practice. However, medical AI training paradigms are shifting towards federated learning settings in order to comply with data protection regulations. In a federated scenario, past data is inaccessible to the current user. Thus, we use a deep generative model to generate synthetic examples that protect privacy and explain decisions. Our proof-of-concept focuses on pleural effusion diagnosis and uses publicly available Chest X-ray data.

Valentina Corbetta, Floris Six Dijkstra, Regina Beets-Tan et al.

Deep learning models in medical imaging often achieve strong in-distribution performance but struggle to generalise under distribution shifts, frequently relying on spurious correlations instead of clinically meaningful features. We introduce LCRReg, a novel regularisation approach that leverages Latent Concept Representations (LCRs) (e.g., Concept Activation Vectors (CAVs)) to guide models toward semantically grounded representations. LCRReg requires no concept labels in the main training set and instead uses a small auxiliary dataset to synthesise high-quality, disentangled concept examples. We extract LCRs for predefined relevant features, and incorporate a regularisation term that guides a Convolutional Neural Network (CNN) to activate within latent subspaces associated with those concepts. We evaluate LCRReg across synthetic and real-world medical tasks. On a controlled toy dataset, it significantly improves robustness to injected spurious correlations and remains effective even in multi-concept and multiclass settings. On the diabetic retinopathy binary classification task, LCRReg enhances performance under both synthetic spurious perturbations and out-of-distribution (OOD) generalisation. Compared to baselines, including multitask learning, linear probing, and post-hoc concept-based models, LCRReg offers a lightweight, architecture-agnostic strategy for improving model robustness without requiring dense concept supervision. Code is available at the following link: https://github.com/Trustworthy-AI-UU-NKI/lcr\_regularization

Aniek Eijpe, Soufyan Lakbir, Melis Erdal Cesur et al.

To improve the prediction of cancer survival using whole-slide images and transcriptomics data, it is crucial to capture both modality-shared and modality-specific information. However, multimodal frameworks often entangle these representations, limiting interpretability and potentially suppressing discriminative features. To address this, we propose Disentangled and Interpretable Multimodal Attention Fusion (DIMAF), a multimodal framework that separates the intra- and inter-modal interactions within an attention-based fusion mechanism to learn distinct modality-specific and modality-shared representations. We introduce a loss based on Distance Correlation to promote disentanglement between these representations and integrate Shapley additive explanations to assess their relative contributions to survival prediction. We evaluate DIMAF on four public cancer survival datasets, achieving a relative average improvement of 1.85% in performance and 23.7% in disentanglement compared to current state-of-the-art multimodal models. Beyond improved performance, our interpretable framework enables a deeper exploration of the underlying interactions between and within modalities in cancer biology.

Miriam Cobo, David Corral Fontecha, Wilson Silva et al.

Artificial intelligence in medical imaging has seen unprecedented growth in the last years, due to rapid advances in deep learning and computing resources. Applications cover the full range of existing medical imaging modalities, with unique characteristics driven by the physics of each technique. Yet, artificial intelligence professionals entering the field, and even experienced developers, often lack a comprehensive understanding of the physical principles underlying medical image acquisition, which hinders their ability to fully leverage its potential. The integration of physics knowledge into artificial intelligence algorithms enhances their trustworthiness and robustness in medical imaging, especially in scenarios with limited data availability. In this work, we review the fundamentals of physics in medical images and their impact on the latest advances in artificial intelligence, particularly, in generative models and reconstruction algorithms. Finally, we explore the integration of physics knowledge into physics-inspired machine learning models, which leverage physics-based constraints to enhance the learning of medical imaging features.

Carla Monteiro, Valentina Corbetta, Regina Beets-Tan et al.

Automatic polyp segmentation is crucial for improving the clinical identification of colorectal cancer (CRC). While Deep Learning (DL) techniques have been extensively researched for this problem, current methods frequently struggle with generalization, particularly in data-constrained or challenging settings. Moreover, many existing polyp segmentation methods rely on complex, task-specific architectures. To address these limitations, we present a framework that leverages the intrinsic robustness of DINO self-attention "key" features for robust segmentation. Unlike traditional methods that extract tokens from the deepest layers of the Vision Transformer (ViT), our approach leverages the key features of the self-attention module with a simple convolutional decoder to predict polyp masks, resulting in enhanced performance and better generalizability. We validate our approach using a multi-center dataset under two rigorous protocols: Domain Generalization (DG) and Extreme Single Domain Generalization (ESDG). Our results, supported by a comprehensive statistical analysis, demonstrate that this pipeline achieves state-of-the-art (SOTA) performance, significantly enhancing generalization, particularly in data-scarce and challenging scenarios. While avoiding a polyp-specific architecture, we surpass well-established models like nnU-Net and UM-Net. Additionally, we provide a systematic benchmark of the DINO framework's evolution, quantifying the specific impact of architectural advancements on downstream polyp segmentation performance.

Daniël Boeke, Cedrik Blommestijn, Rebecca N. Wray et al.

Recurrence risk estimation in clear cell renal cell carcinoma (ccRCC) is essential for guiding postoperative surveillance and treatment. The Leibovich score remains widely used for stratifying distant recurrence risk but offers limited patient-level resolution and excludes imaging information. This study evaluates multimodal recurrence prediction by integrating preoperative computed tomography (CT) and postoperative histopathology whole-slide images (WSIs). A modular deep learning framework with pretrained encoders and Cox-based survival modeling was tested across unimodal, late fusion, and intermediate fusion setups. In a real-world ccRCC cohort, WSI-based models consistently outperformed CT-only models, underscoring the prognostic strength of pathology. Intermediate fusion further improved performance, with the best model (TITAN-CONCH with ResNet-18) approaching the adjusted Leibovich score. Random tie-breaking narrowed the gap between the clinical baseline and learned models, suggesting discretization may overstate individualized performance. Using simple embedding concatenation, radiology added value primarily through fusion. These findings demonstrate the feasibility of foundation model-based multimodal integration for personalized ccRCC risk prediction. Future work should explore more expressive fusion strategies, larger multimodal datasets, and general-purpose CT encoders to better match pathology modeling capacity.

Gideon N. L. Rouwendaal, Daniël Boeke, Inge L. Cox et al.

Accurate preoperative prediction of erectile dysfunction (ED) is important for counseling patients undergoing radical prostatectomy. While clinical features are established predictors, the added value of preoperative MRI remains underexplored. We investigate whether MRI provides additional predictive value for ED at 12 months post-surgery, evaluating four modeling strategies: (1) a clinical-only baseline, representing current state-of-the-art; (2) classical models using handcrafted anatomical features derived from MRI; (3) deep learning models trained directly on MRI slices; and (4) multimodal fusion of imaging and clinical inputs. Imaging-based models (maximum AUC 0.569) slightly outperformed handcrafted anatomical approaches (AUC 0.554) but fell short of the clinical baseline (AUC 0.663). Fusion models offered marginal gains (AUC 0.586) but did not exceed clinical-only performance. SHAP analysis confirmed that clinical features contributed most to predictive performance. Saliency maps from the best-performing imaging model suggested a predominant focus on anatomically plausible regions, such as the prostate and neurovascular bundles. While MRI-based models did not improve predictive performance over clinical features, our findings suggest that they try to capture patterns in relevant anatomical structures and may complement clinical predictors in future multimodal approaches.

Jan van Eck, Dea Gogishvili, Wilson Silva et al.

Protein language models (PLMs) have revolutionised computational biology through their ability to generate powerful sequence representations for diverse prediction tasks. However, their black-box nature limits biological interpretation and translation to actionable insights. We present an explainable adapter layer - PLM-eXplain (PLM-X), that bridges this gap by factoring PLM embeddings into two components: an interpretable subspace based on established biochemical features, and a residual subspace that preserves the model's predictive power. Using embeddings from ESM2, our adapter incorporates well-established properties, including secondary structure and hydropathy while maintaining high performance. We demonstrate the effectiveness of our approach across three protein-level classification tasks: prediction of extracellular vesicle association, identification of transmembrane helices, and prediction of aggregation propensity. PLM-X enables biological interpretation of model decisions without sacrificing accuracy, offering a generalisable solution for enhancing PLM interpretability across various downstream applications. This work addresses a critical need in computational biology by providing a bridge between powerful deep learning models and actionable biological insights.