Enzo Ferrante

Karim Lekadir, Aasa Feragen, Abdul Joseph Fofanah et al. · eth-zurich

Despite major advances in artificial intelligence (AI) for medicine and healthcare, the deployment and adoption of AI technologies remain limited in real-world clinical practice. In recent years, concerns have been raised about the technical, clinical, ethical and legal risks associated with medical AI. To increase real world adoption, it is essential that medical AI tools are trusted and accepted by patients, clinicians, health organisations and authorities. This work describes the FUTURE-AI guideline as the first international consensus framework for guiding the development and deployment of trustworthy AI tools in healthcare. The FUTURE-AI consortium was founded in 2021 and currently comprises 118 inter-disciplinary experts from 51 countries representing all continents, including AI scientists, clinicians, ethicists, and social scientists. Over a two-year period, the consortium defined guiding principles and best practices for trustworthy AI through an iterative process comprising an in-depth literature review, a modified Delphi survey, and online consensus meetings. The FUTURE-AI framework was established based on 6 guiding principles for trustworthy AI in healthcare, i.e. Fairness, Universality, Traceability, Usability, Robustness and Explainability. Through consensus, a set of 28 best practices were defined, addressing technical, clinical, legal and socio-ethical dimensions. The recommendations cover the entire lifecycle of medical AI, from design, development and validation to regulation, deployment, and monitoring. FUTURE-AI is a risk-informed, assumption-free guideline which provides a structured approach for constructing medical AI tools that will be trusted, deployed and adopted in real-world practice. Researchers are encouraged to take the recommendations into account in proof-of-concept stages to facilitate future translation towards clinical practice of medical AI.

Leo Fillioux, Enzo Ferrante, Paul-Henry Cournède et al.

Large foundation models have emerged in the last years and are pushing performance boundaries for a variety of tasks. Training or even finetuning such models demands vast datasets and computational resources, which are often scarce and costly. Adaptation methods provide a computationally efficient solution to address these limitations by allowing such models to be finetuned on small amounts of data and computing power. This is achieved by appending new trainable modules to frozen backbones with only a fraction of the trainable parameters and fitting only these modules on novel tasks. Recently, the VeRA adapter was shown to excel in parameter-efficient adaptations by utilizing a pair of frozen random low-rank matrices shared across all layers. In this paper, we propose PVeRA, a probabilistic version of the VeRA adapter, which modifies the low-rank matrices of VeRA in a probabilistic manner. This modification naturally allows handling inherent ambiguities in the input and allows for different sampling configurations during training and testing. A comprehensive evaluation was performed on the VTAB-1k benchmark and seven adapters, with PVeRA outperforming VeRA and other adapters. Our code for training models with PVeRA and benchmarking all adapters is available https://github.com/leofillioux/pvera.

Nicolás Gaggion, Lucas Mansilla, Candelaria Mosquera et al.

Anatomical segmentation is a fundamental task in medical image computing, generally tackled with fully convolutional neural networks which produce dense segmentation masks. These models are often trained with loss functions such as cross-entropy or Dice, which assume pixels to be independent of each other, thus ignoring topological errors and anatomical inconsistencies. We address this limitation by moving from pixel-level to graph representations, which allow to naturally incorporate anatomical constraints by construction. To this end, we introduce HybridGNet, an encoder-decoder neural architecture that leverages standard convolutions for image feature encoding and graph convolutional neural networks (GCNNs) to decode plausible representations of anatomical structures. We also propose a novel image-to-graph skip connection layer which allows localized features to flow from standard convolutional blocks to GCNN blocks, and show that it improves segmentation accuracy. The proposed architecture is extensively evaluated in a variety of domain shift and image occlusion scenarios, and audited considering different types of demographic domain shift. Our comprehensive experimental setup compares HybridGNet with other landmark and pixel-based models for anatomical segmentation in chest x-ray images, and shows that it produces anatomically plausible results in challenging scenarios where other models tend to fail.

Nicolás Gaggion, Candelaria Mosquera, Lucas Mansilla et al.

The development of successful artificial intelligence models for chest X-ray analysis relies on large, diverse datasets with high-quality annotations. While several databases of chest X-ray images have been released, most include disease diagnosis labels but lack detailed pixel-level anatomical segmentation labels. To address this gap, we introduce an extensive chest X-ray multi-center segmentation dataset with uniform and fine-grain anatomical annotations for images coming from five well-known publicly available databases: ChestX-ray8, Chexpert, MIMIC-CXR-JPG, Padchest, and VinDr-CXR, resulting in 657,566 segmentation masks. Our methodology utilizes the HybridGNet model to ensure consistent and high-quality segmentations across all datasets. Rigorous validation, including expert physician evaluation and automatic quality control, was conducted to validate the resulting masks. Additionally, we provide individualized quality indices per mask and an overall quality estimation per dataset. This dataset serves as a valuable resource for the broader scientific community, streamlining the development and assessment of innovative methodologies in chest X-ray analysis. The CheXmask dataset is publicly available at: https://physionet.org/content/chexmask-cxr-segmentation-data/

Rodrigo Bonazzola, Enzo Ferrante, Nishant Ravikumar et al.

Recent genome-wide association studies (GWAS) have been successful in identifying associations between genetic variants and simple cardiac parameters derived from cardiac magnetic resonance (CMR) images. However, the emergence of big databases including genetic data linked to CMR, facilitates investigation of more nuanced patterns of shape variability. Here, we propose a new framework for gene discovery entitled Unsupervised Phenotype Ensembles (UPE). UPE builds a redundant yet highly expressive representation by pooling a set of phenotypes learned in an unsupervised manner, using deep learning models trained with different hyperparameters. These phenotypes are then analyzed via (GWAS), retaining only highly confident and stable associations across the ensemble. We apply our approach to the UK Biobank database to extract left-ventricular (LV) geometric features from image-derived three-dimensional meshes. We demonstrate that our approach greatly improves the discoverability of genes influencing LV shape, identifying 11 loci with study-wide significance and 8 with suggestive significance. We argue that our approach would enable more extensive discovery of gene associations with image-derived phenotypes for other organs or image modalities.

Nicolás Gaggion, Benjamin A. Matheson, Yan Xia et al.

Cardiovascular magnetic resonance imaging is emerging as a crucial tool to examine cardiac morphology and function. Essential to this endeavour are anatomical 3D surface and volumetric meshes derived from CMR images, which facilitate computational anatomy studies, biomarker discovery, and in-silico simulations. Traditional approaches typically follow complex multi-step pipelines, first segmenting images and then reconstructing meshes, making them time-consuming and prone to error propagation. In response, we introduce HybridVNet, a novel architecture for direct image-to-mesh extraction seamlessly integrating standard convolutional neural networks with graph convolutions, which we prove can efficiently handle surface and volumetric meshes by encoding them as graph structures. To further enhance accuracy, we propose a multi-view HybridVNet architecture which processes both long axis and short axis CMR, showing that it can increase the performance of cardiac MR mesh generation. Our model combines traditional convolutional networks with variational graph generative models, deep supervision and mesh-specific regularisation. Experiments on a comprehensive dataset from the UK Biobank confirm the potential of HybridVNet to significantly advance cardiac imaging and computational cardiology by efficiently generating high-fidelity meshes from CMR images. Multi-view HybridVNet outperforms the state-of-the-art, achieving improvements of up to $\sim$27\% reduction in Mean Contour Distance (from 1.86 mm to 1.35 mm for the LV Myocardium), up to $\sim$18\% improvement in Hausdorff distance (from 4.74 mm to 3.89mm, for the LV Endocardium), and up to $\sim$8\% in Dice Coefficient (from 0.78 to 0.84, for the LV Myocardium), highlighting its superior accuracy.

Nicolás Gaggion, Maria Vakalopoulou, Diego H. Milone et al.

Learning anatomical segmentation from heterogeneous labels in multi-center datasets is a common situation encountered in clinical scenarios, where certain anatomical structures are only annotated in images coming from particular medical centers, but not in the full database. Here we first show how state-of-the-art pixel-level segmentation models fail in naively learning this task due to domain memorization issues and conflicting labels. We then propose to adopt HybridGNet, a landmark-based segmentation model which learns the available anatomical structures using graph-based representations. By analyzing the latent space learned by both models, we show that HybridGNet naturally learns more domain-invariant feature representations, and provide empirical evidence in the context of chest X-ray multiclass segmentation. We hope these insights will shed light on the training of deep learning models with heterogeneous labels from public and multi-center datasets.

Badr-Eddine Marani, Mohamed Hanini, Nihitha Malayarukil et al.

The proliferation of machine learning models in critical decision making processes has underscored the need for bias discovery and mitigation strategies. Identifying the reasons behind a biased system is not straightforward, since in many occasions they are associated with hidden spurious correlations which are not easy to spot. Standard approaches rely on bias audits performed by analyzing model performance in pre-defined subgroups of data samples, usually characterized by common attributes like gender or ethnicity when it comes to people, or other specific attributes defining semantically coherent groups of images. However, it is not always possible to know a-priori the specific attributes defining the failure modes of visual recognition systems. Recent approaches propose to discover these groups by leveraging large vision language models, which enable the extraction of cross-modal embeddings and the generation of textual descriptions to characterize the subgroups where a certain model is underperforming. In this work, we argue that incorporating visual explanations (e.g. heatmaps generated via GradCAM or other approaches) can boost the performance of such bias discovery and mitigation frameworks. To this end, we introduce Visually Grounded Bias Discovery and Mitigation (ViG-Bias), a simple yet effective technique which can be integrated to a variety of existing frameworks to improve both, discovery and mitigation performance. Our comprehensive evaluation shows that incorporating visual explanations enhances existing techniques like DOMINO, FACTS and Bias-to-Text, across several challenging datasets, including CelebA, Waterbirds, and NICO++.

Nicolás Gaggion, Maria J. Ledesma-Carbayo, Stergios Christodoulidis et al.

Graph-based medical image segmentation represents anatomical structures using boundary graphs, providing fixed-topology landmarks and inherent population-level correspondences. However, their clinical adoption has been hindered by a major requirement: training datasets with manually annotated landmarks that maintain point-to-point correspondences across patients rarely exist in practice. We introduce Mask-HybridGNet, a framework that trains graph-based models directly using standard pixel-wise masks, eliminating the need for manual landmark annotations. Our approach aligns variable-length ground truth boundaries with fixed-length landmark predictions by combining Chamfer distance supervision and edge-based regularization to ensure local smoothness and regular landmark distribution, further refined via differentiable rasterization. A significant emergent property of this framework is that predicted landmark positions become consistently associated with specific anatomical locations across patients without explicit correspondence supervision. This implicit atlas learning enables temporal tracking, cross-slice reconstruction, and morphological population analyses. Beyond direct segmentation, Mask-HybridGNet can extract correspondences from existing segmentation masks, allowing it to generate stable anatomical atlases from any high-quality pixel-based model. Experiments across chest radiography, cardiac ultrasound, cardiac MRI, and fetal imaging demonstrate that our model achieves competitive results against state-of-the-art pixel-based methods, while ensuring anatomical plausibility by enforcing boundary connectivity through a fixed graph adjacency matrix. This framework leverages the vast availability of standard segmentation masks to build structured models that maintain topological integrity and provide implicit correspondences.

Amine Sadikine, Bogdan Badic, Enzo Ferrante et al.

The precise delineation of blood vessels in medical images is critical for many clinical applications, including pathology detection and surgical planning. However, fully-automated vascular segmentation is challenging because of the variability in shape, size, and topology. Manual segmentation remains the gold standard but is time-consuming, subjective, and impractical for large-scale studies. Hence, there is a need for automatic and reliable segmentation methods that can accurately detect blood vessels from medical images. The integration of shape and topological priors into vessel segmentation models has been shown to improve segmentation accuracy by offering contextual information about the shape of the blood vessels and their spatial relationships within the vascular tree. To further improve anatomical consistency, we propose a new joint prior encoding mechanism which incorporates both shape and topology in a single latent space. The effectiveness of our method is demonstrated on the publicly available 3D-IRCADb dataset. More globally, the proposed approach holds promise in overcoming the challenges associated with automatic vessel delineation and has the potential to advance the field of deep priors encoding.

Dovile Juodelyte, Yucheng Lu, Amelia Jiménez-Sánchez et al.

Transfer learning has become an essential part of medical imaging classification algorithms, often leveraging ImageNet weights. The domain shift from natural to medical images has prompted alternatives such as RadImageNet, often showing comparable classification performance. However, it remains unclear whether the performance gains from transfer learning stem from improved generalization or shortcut learning. To address this, we conceptualize confounders by introducing the Medical Imaging Contextualized Confounder Taxonomy (MICCAT) and investigate a range of confounders across it -- whether synthetic or sampled from the data -- using two public chest X-ray and CT datasets. We show that ImageNet and RadImageNet achieve comparable classification performance, yet ImageNet is much more prone to overfitting to confounders. We recommend that researchers using ImageNet-pretrained models reexamine their model robustness by conducting similar experiments. Our code and experiments are available at https://github.com/DovileDo/source-matters.

Nicolás Gaggion, Enzo Ferrante, Beatriz Paniagua et al.

Skeletonization is a popular shape analysis technique that models an object's interior as opposed to just its boundary. Fitting template-based skeletal models is a time-consuming process requiring much manual parameter tuning. Recently, machine learning-based methods have shown promise for generating s-reps from object boundaries. In this work, we propose a new skeletonization method which leverages graph convolutional networks to produce skeletal representations (s-reps) from dense segmentation masks. The method is evaluated on both synthetic data and real hippocampus segmentations, achieving promising results and fast inference.

Enzo Ferrante, Rodrigo Echeveste

Recently, the research community of computerized medical imaging has started to discuss and address potential fairness issues that may emerge when developing and deploying AI systems for medical image analysis. This chapter covers some of the pressing challenges encountered when doing research in this area, and it is intended to raise questions and provide food for thought for those aiming to enter this research field. The chapter first discusses various sources of bias, including data collection, model training, and clinical deployment, and their impact on the fairness of machine learning algorithms in medical image computing. We then turn to discussing open challenges that we believe require attention from researchers and practitioners, as well as potential pitfalls of naive application of common methods in the field. We cover a variety of topics including the impact of biased metrics when auditing for fairness, the leveling down effect, task difficulty variations among subgroups, discovering biases in unseen populations, and explaining biases beyond standard demographic attributes.

David Restrepo, Ira Ktena, Maria Vakalopoulou et al.

Vision-language models (VLMs) are increasingly proposed for clinical decision support, yet their reliability in real-world scenarios that require integrating both visual and textual context from medical records remains poorly characterized. This paper identifies three failure modes: (1) modality over-reliance on text over images, (2) spurious reliance on irrelevant clinical history, and (3) prompt sensitivity across semantically equivalent inputs. We evaluate a diverse set of general-domain and medically-tuned open and closed VLMs on chest x-ray tasks using MIMIC-CXR. By systematically manipulating image-text alignment, clinical history, and prompt formulations, we found that VLM decisions are dominated by the text modality, even when visual evidence is available. Moreover, we observed that VLMs are heavily influenced by irrelevant reports, while minor prompt changes can reverse correct image-based predictions. Our findings underscore the need for explicit safeguards and stress-testing before considering the use of these models in clinical practice.

David Restrepo, Miguel L Martins, Chenwei Wu et al.

Vision-Language Models (VLMs) exhibit a characteristic "cone effect" in which nonlinear encoders map embeddings into highly concentrated regions of the representation space, contributing to cross-modal separation known as the modality gap. While this phenomenon has been widely observed, its practical impact on supervised multimodal learning -particularly in medical domains- remains unclear. In this work, we introduce a lightweight post-hoc mechanism that keeps pretrained VLM encoders frozen while continuously controlling cross-modal separation through a single hyperparameter {λ}. This enables systematic analysis of how the modality gap affects downstream multimodal performance without expensive retraining. We evaluate generalist (CLIP, SigLIP) and medically specialized (BioMedCLIP, MedSigLIP) models across diverse medical and natural datasets in a supervised multimodal settings. Results consistently show that reducing excessive modality gap improves downstream performance, with medical datasets exhibiting stronger sensitivity to gap modulation; however, fully collapsing the gap is not always optimal, and intermediate, task-dependent separation yields the best results. These findings position the modality gap as a tunable property of multimodal representations rather than a quantity that should be universally minimized.

Dovile Juodelyte, Enzo Ferrante, Yucheng Lu et al.

Current transferability estimation methods designed for natural image datasets are often suboptimal in medical image classification. These methods primarily focus on estimating the suitability of pre-trained source model features for a target dataset, which can lead to unrealistic predictions, such as suggesting that the target dataset is the best source for itself. To address this, we propose a novel transferability metric that combines feature quality with gradients to evaluate both the suitability and adaptability of source model features for target tasks. We evaluate our approach in two new scenarios: source dataset transferability for medical image classification and cross-domain transferability. Our results show that our method outperforms existing transferability metrics in both settings. We also provide insight into the factors influencing transfer performance in medical image classification, as well as the dynamics of cross-domain transfer from natural to medical images. Additionally, we provide ground-truth transfer performance benchmarking results to encourage further research into transferability estimation for medical image classification. Our code and experiments are available at https://github.com/DovileDo/transferability-in-medical-imaging.

Matias Cosarinsky, Ramiro Billot, Lucas Mansilla et al.

Assessing the quality of automatic image segmentation is crucial in clinical practice, but often very challenging due to the limited availability of ground truth annotations. Reverse Classification Accuracy (RCA) is an approach that estimates the quality of new predictions on unseen samples by training a segmenter on those predictions, and then evaluating it against existing annotated images. In this work, we introduce Conformal In-Context RCA, a novel method for automatically estimating segmentation quality with statistical guarantees in the absence of ground-truth annotations, which consists of two main innovations. First, In-Context RCA, which leverages recent in-context learning models for image segmentation and incorporates retrieval-augmentation techniques to select the most relevant reference images. This approach enables efficient quality estimation with minimal reference data while avoiding the need of training additional models. Second, we introduce Conformal RCA, which extends both the original RCA framework and In-Context RCA to go beyond point estimation. Using tools from split conformal prediction, Conformal RCA produces prediction intervals for segmentation quality providing statistical guarantees that the true score lies within the estimated interval with a user-specified probability. Validated across 10 different medical imaging tasks in various organs and modalities, our methods demonstrate robust performance and computational efficiency, offering a promising solution for automated quality control in clinical workflows, where fast and reliable segmentation assessment is essential. The code is available at https://github.com/mcosarinsky/Conformal-In-Context-RCA.

Matias Cosarinsky, Nicolas Gaggion, Rodrigo Echeveste et al.

In this work, we study uncertainty estimation for anatomical landmark-based segmentation on chest X-rays. Inspired by hybrid neural network architectures that combine standard image convolutional encoders with graph-based generative decoders, and leveraging their variational latent space, we derive two complementary measures: (i) latent uncertainty, captured directly from the learned distribution parameters, and (ii) predictive uncertainty, obtained by generating multiple stochastic output predictions from latent samples. Through controlled corruption experiments we show that both uncertainty measures increase with perturbation severity, reflecting both global and local degradation. We demonstrate that these uncertainty signals can identify unreliable predictions by comparing with manual ground-truth, and support out-of-distribution detection on the CheXmask dataset. More importantly, we release CheXmask-U (huggingface.co/datasets/mcosarinsky/CheXmask-U), a large scale dataset of 657,566 chest X-ray landmark segmentations with per-node uncertainty estimates, enabling researchers to account for spatial variations in segmentation quality when using these anatomical masks. Our findings establish uncertainty estimation as a promising direction to enhance robustness and safe deployment of landmark-based anatomical segmentation methods in chest X-ray. A fully working interactive demo of the method is available at huggingface.co/spaces/matiasky/CheXmask-U and the source code at github.com/mcosarinsky/CheXmask-U.

David Restrepo, Ira Ktena, Maria Vakalopoulou et al.

Clinical decision-making relies on the integrated analysis of medical images and the associated clinical reports. While Vision-Language Models (VLMs) can offer a unified framework for such tasks, they can exhibit strong biases toward one modality, frequently overlooking critical visual cues in favor of textual information. In this work, we introduce Selective Modality Shifting (SMS), a perturbation-based approach to quantify a model's reliance on each modality in binary classification tasks. By systematically swapping images or text between samples with opposing labels, we expose modality-specific biases. We assess six open-source VLMs-four generalist models and two fine-tuned for medical data-on two medical imaging datasets with distinct modalities: MIMIC-CXR (chest X-ray) and FairVLMed (scanning laser ophthalmoscopy). By assessing model performance and the calibration of every model in both unperturbed and perturbed settings, we reveal a marked dependency on text input, which persists despite the presence of complementary visual information. We also perform a qualitative attention-based analysis which further confirms that image content is often overshadowed by text details. Our findings highlight the importance of designing and evaluating multimodal medical models that genuinely integrate visual and textual cues, rather than relying on single-modality signals.

Théo Sourget, David Restrepo, Céline Hudelot et al.

Motivated by the strong performance of CLIP-based models in natural image-text domains, recent efforts have adapted these architectures to medical tasks, particularly in radiology, where large paired datasets of images and reports, such as chest X-rays, are available. While these models have shown encouraging results in terms of accuracy and discriminative performance, their fairness and robustness in the different clinical tasks remain largely underexplored. In this study, we extensively evaluate six widely used CLIP-based models on chest X-ray classification using three publicly available datasets: MIMIC-CXR, NIH-CXR14, and NEATX. We assess the models fairness across six conditions and patient subgroups based on age, sex, and race. Additionally, we assess the robustness to shortcut learning by evaluating performance on pneumothorax cases with and without chest drains. Our results indicate performance gaps between patients of different ages, but more equitable results for the other attributes. Moreover, all models exhibit lower performance on images without chest drains, suggesting reliance on spurious correlations. We further complement the performance analysis with a study of the embeddings generated by the models. While the sensitive attributes could be classified from the embeddings, we do not see such patterns using PCA, showing the limitations of these visualisation techniques when assessing models. Our code is available at https://github.com/TheoSourget/clip_cxr_fairness

Nicolás Gaggion, Rodrigo Bonazzola, María Florencia Legascue et al.

The analysis of plant developmental plasticity, including root system architecture, is fundamental to understanding plant adaptability and development, particularly in the context of climate change and agricultural sustainability. While significant advances have been made in plant phenotyping technologies, comprehensive temporal analysis of root development remains challenging, with most existing solutions providing either limited throughput or restricted structural analysis capabilities. Here, we present ChronoRoot 2.0, an integrated open-source platform that combines affordable hardware with advanced artificial intelligence to enable sophisticated temporal plant phenotyping. The system introduces several major advances, offering an integral perspective of seedling development: (i) simultaneous multi-organ tracking of six distinct plant structures, (ii) quality control through real-time validation, (iii) comprehensive architectural measurements including novel gravitropic response parameters, and (iv) dual specialized user interfaces for both architectural analysis and high-throughput screening. We demonstrate the system's capabilities through three use cases for Arabidopsis thaliana: characterization of circadian growth patterns under different light conditions, detailed analysis of gravitropic responses in transgenic plants, and high-throughput screening of etiolation responses across multiple genotypes. ChronoRoot 2.0 maintains its predecessor's advantages of low cost and modularity while significantly expanding its capabilities, making sophisticated temporal phenotyping more accessible to the broader plant science community. The system's open-source nature, combined with extensive documentation and containerized deployment options, ensures reproducibility and enables community-driven development of new analytical capabilities.

José Ignacio Orlando, Hugo Luis Manterola, Enzo Ferrante et al.

Arabidopsis thaliana is a plant species widely utilized by scientists to estimate the impact of genetic differences in root morphological features. For this purpose, images of this plant after genetic modifications are taken to study differences in the root architecture. This task requires manual segmentations of radicular structures, although this is a particularly tedious and time-consuming labor. In this work, we present an unsupervised method for Arabidopsis thaliana root segmentation based on morphological operations and fully-connected Conditional Random Fields. Although other approaches have been proposed to this purpose, all of them are based on more complex and expensive imaging modalities. Our results prove that our method can be easily applied over images taken using conventional scanners, with a minor user intervention. A first data set, our results and a fully open source implementation are available online.

Shivalika Singh, Angelika Romanou, Clémentine Fourrier et al.

Cultural biases in multilingual datasets pose significant challenges for their effectiveness as global benchmarks. These biases stem not only from differences in language but also from the cultural knowledge required to interpret questions, reducing the practical utility of translated datasets like MMLU. Furthermore, translation often introduces artefacts that can distort the meaning or clarity of questions in the target language. A common practice in multilingual evaluation is to rely on machine-translated evaluation sets, but simply translating a dataset is insufficient to address these challenges. In this work, we trace the impact of both of these issues on multilingual evaluations and ensuing model performances. Our large-scale evaluation of state-of-the-art open and proprietary models illustrates that progress on MMLU depends heavily on learning Western-centric concepts, with 28% of all questions requiring culturally sensitive knowledge. Moreover, for questions requiring geographic knowledge, an astounding 84.9% focus on either North American or European regions. Rankings of model evaluations change depending on whether they are evaluated on the full portion or the subset of questions annotated as culturally sensitive, showing the distortion to model rankings when blindly relying on translated MMLU. We release Global MMLU, an improved MMLU with evaluation coverage across 42 languages -- with improved overall quality by engaging with compensated professional and community annotators to verify translation quality while also rigorously evaluating cultural biases present in the original dataset. This comprehensive Global MMLU set also includes designated subsets labeled as culturally sensitive and culturally agnostic to allow for more holistic, complete evaluation.

Manuel Fernández Burda, Santiago Aranguri, Iván Arcuschin Moreno et al.

Protein language models (PLMs) are becoming practical tools for de novo protein design, yet their dual-use potential raises safety concerns. We show that domain adaptation to specific taxonomic groups can elicit toxic protein generation, even when toxicity is not the training objective. To address this, we adapt Logit Diff Amplification (LDA) as an inference-time control mechanism for PLMs. LDA modifies token probabilities by amplifying the logit difference between a baseline model and a toxicity-finetuned model, requiring no retraining. Across four taxonomic groups, LDA consistently reduces predicted toxicity rate (measured via ToxDL2) below the taxon-finetuned baseline while preserving biological plausibility. We evaluate quality using Fréchet ESM Distance and predicted foldability (pLDDT), finding that LDA maintains distributional similarity to natural proteins and structural viability (unlike activation-based steering methods that tend to degrade sequence properties). Our results demonstrate that LDA provides a practical safety knob for protein generators that mitigates elicited toxicity while retaining generative quality.

Amelia Jiménez-Sánchez, Natalia-Rozalia Avlona, Sarah de Boer et al.

Datasets play a critical role in medical imaging research, yet issues such as label quality, shortcuts, and metadata are often overlooked. This lack of attention may harm the generalizability of algorithms and, consequently, negatively impact patient outcomes. While existing medical imaging literature reviews mostly focus on machine learning (ML) methods, with only a few focusing on datasets for specific applications, these reviews remain static -- they are published once and not updated thereafter. This fails to account for emerging evidence, such as biases, shortcuts, and additional annotations that other researchers may contribute after the dataset is published. We refer to these newly discovered findings of datasets as research artifacts. To address this gap, we propose a living review that continuously tracks public datasets and their associated research artifacts across multiple medical imaging applications. Our approach includes a framework for the living review to monitor data documentation artifacts, and an SQL database to visualize the citation relationships between research artifact and dataset. Lastly, we discuss key considerations for creating medical imaging datasets, review best practices for data annotation, discuss the significance of shortcuts and demographic diversity, and emphasize the importance of managing datasets throughout their entire lifecycle. Our demo is publicly available at http://inthepicture.itu.dk/.

Cynthia Maldonado-Garcia, Rodrigo Bonazzola, Enzo Ferrante et al.

Cardiovascular diseases (CVD) are the leading cause of death globally. Non-invasive, cost-effective imaging techniques play a crucial role in early detection and prevention of CVD. Optical coherence tomography (OCT) has gained recognition as a potential tool for early CVD risk prediction, though its use remains underexplored. In this study, we investigated the potential of OCT as an additional imaging technique to predict future CVD events. We analysed retinal OCT data from the UK Biobank. The dataset included 612 patients who suffered a myocardial infarction (MI) or stroke within five years of imaging and 2,234 controls without CVD (total: 2,846 participants). A self-supervised deep learning approach based on Variational Autoencoders (VAE) was used to extract low-dimensional latent representations from high-dimensional 3D OCT images, capturing distinct features of retinal layers. These latent features, along with clinical data, were used to train a Random Forest (RF) classifier to differentiate between patients at risk of future CVD events (MI or stroke) and healthy controls. Our model achieved an AUC of 0.75, sensitivity of 0.70, specificity of 0.70, and accuracy of 0.70, outperforming the QRISK3 score (the third version of the QRISK cardiovascular disease risk prediction algorithm; AUC = 0.60, sensitivity = 0.60, specificity = 0.55, accuracy = 0.55). The choroidal layer in OCT images was identified as a key predictor of future CVD events, revealed through a novel model explainability approach. This study demonstrates that retinal OCT imaging is a cost-effective, non-invasive alternative for predicting CVD risk, offering potential for widespread application in optometry practices and hospitals.

Israfel Salazar, Manuel Fernández Burda, Shayekh Bin Islam et al. · mila

The evaluation of vision-language models (VLMs) has mainly relied on English-language benchmarks, leaving significant gaps in both multilingual and multicultural coverage. While multilingual benchmarks have expanded, both in size and languages, many rely on translations of English datasets, failing to capture cultural nuances. In this work, we propose Kaleidoscope, as the most comprehensive exam benchmark to date for the multilingual evaluation of vision-language models. Kaleidoscope is a large-scale, in-language multimodal benchmark designed to evaluate VLMs across diverse languages and visual inputs. Kaleidoscope covers 18 languages and 14 different subjects, amounting to a total of 20,911 multiple-choice questions. Built through an open science collaboration with a diverse group of researchers worldwide, Kaleidoscope ensures linguistic and cultural authenticity. We evaluate top-performing multilingual vision-language models and find that they perform poorly on low-resource languages and in complex multimodal scenarios. Our results highlight the need for progress on culturally inclusive multimodal evaluation frameworks.

Franco Matzkin, Agostina Larrazabal, Diego H Milone et al.

Accurate segmentation of white matter hyperintensities (WMH) is crucial for clinical decision-making, particularly in the context of multiple sclerosis. However, domain shifts, such as variations in MRI machine types or acquisition parameters, pose significant challenges to model calibration and uncertainty estimation. This study investigates the impact of domain shift on WMH segmentation by proposing maximum-entropy regularization techniques to enhance model calibration and uncertainty estimation, with the purpose of identifying errors post-deployment using predictive uncertainty as a proxy measure that does not require ground-truth labels. To do this, we conducted experiments using a U-Net architecture to evaluate these regularization schemes on two publicly available datasets, assessing performance with the Dice coefficient, expected calibration error, and entropy-based uncertainty estimates. Our results show that entropy-based uncertainty estimates can anticipate segmentation errors, and that maximum-entropy regularization further strengthens the correlation between uncertainty and segmentation performance while also improving model calibration under domain shift.

Josefina Catoni, Domonkos Martos, Ferenc Csikor et al.

Deep Generative Models (DGMs) can learn flexible latent variable representations of images while avoiding intractable computations, common in Bayesian inference. However, investigating the properties of inference in Variational Autoencoders (VAEs), a major class of DGMs, reveals severe problems in their uncertainty representations. Here we draw inspiration from classical computer vision to introduce an inductive bias into the VAE by incorporating a global explaining-away latent variable, which remedies defective inference in VAEs. Unlike standard VAEs, the Explaing-Away VAE (EA-VAE) provides uncertainty estimates that align with normative requirements across a wide spectrum of perceptual tasks, including image corruption, interpolation, and out-of-distribution detection. We find that restored inference capabilities are delivered by developing a motif in the inference network (the encoder) which is widespread in biological neural networks: divisive normalization. Our results establish EA-VAEs as reliable tools to perform inference under deep generative models with appropriate estimates of uncertainty.

Lucas Mansilla, Rodrigo Echeveste, Camila Gonzalez et al.

Biases in machine learning pose significant challenges, particularly when models amplify disparities that affect disadvantaged groups. Traditional bias mitigation techniques often lead to a {\itshape leveling-down effect}, whereby improving outcomes of disadvantaged groups comes at the expense of reduced performance for advantaged groups. This study introduces Bias Mitigation through Continual Learning (BM-CL), a novel framework that leverages the principles of continual learning to address this trade-off. We postulate that mitigating bias is conceptually similar to domain-incremental continual learning, where the model must adjust to changing fairness conditions, improving outcomes for disadvantaged groups without forgetting the knowledge that benefits advantaged groups. Drawing inspiration from techniques such as Learning without Forgetting and Elastic Weight Consolidation, we reinterpret bias mitigation as a continual learning problem. This perspective allows models to incrementally balance fairness objectives, enhancing outcomes for disadvantaged groups while preserving performance for advantaged groups. Experiments on synthetic and real-world image datasets, characterized by diverse sources of bias, demonstrate that the proposed framework mitigates biases while minimizing the loss of original knowledge. Our approach bridges the fields of fairness and continual learning, offering a promising pathway for developing machine learning systems that are both equitable and effective.

Jingchen Zou, Jianqiang Li, Gabriel Jimenez et al.

The performance of medical image segmentation models is usually evaluated using metrics like the Dice score and Hausdorff distance, which compare predicted masks to ground truth annotations. However, when applying the model to unseen data, such as in clinical settings, it is often impractical to annotate all the data, making the model's performance uncertain. To address this challenge, we propose the Segmentation Performance Evaluator (SPE), a framework for estimating segmentation models' performance on unlabeled data. This framework is adaptable to various evaluation metrics and model architectures. Experiments on six publicly available datasets across six evaluation metrics including pixel-based metrics such as Dice score and distance-based metrics like HD95, demonstrated the versatility and effectiveness of our approach, achieving a high correlation (0.956$\pm$0.046) and low MAE (0.025$\pm$0.019) compare with real Dice score on the independent test set. These results highlight its ability to reliably estimate model performance without requiring annotations. The SPE framework integrates seamlessly into any model training process without adding training overhead, enabling performance estimation and facilitating the real-world application of medical image segmentation algorithms. The source code is publicly available

Estanislao Claucich, Sara Hooker, Diego H. Milone et al.

Ensembling is commonly regarded as an effective way to improve the general performance of models in machine learning, while also increasing the robustness of predictions. When it comes to algorithmic fairness, heterogeneous ensembles, composed of multiple model types, have been employed to mitigate biases in terms of demographic attributes such as sex, age or ethnicity. Moreover, recent work has shown how in multi-class problems even simple homogeneous ensembles may favor performance of the worst-performing target classes. While homogeneous ensembles are simpler to implement in practice, it is not yet clear whether their benefits translate to groups defined not in terms of their target class, but in terms of demographic or protected attributes, hence improving fairness. In this work we show how this simple and straightforward method is indeed able to mitigate disparities, particularly benefiting under-performing subgroups. Interestingly, this can be achieved without sacrificing overall performance, which is a common trade-off observed in bias mitigation strategies. Moreover, we analyzed the interplay between two factors which may result in biases: sub-group under-representation and the inherent difficulty of the task for each group. These results revealed that, contrary to popular assumptions, having balanced datasets may be suboptimal if the task difficulty varies between subgroups. Indeed, we found that a perfectly balanced dataset may hurt both the overall performance and the gap between groups. This highlights the importance of considering the interaction between multiple forces at play in fairness.

Pedro Alejandro Dal Bianco, Oscar Agustín Stanchi, Facundo Manuel Quiroga et al.

This paper presents the first comprehensive interpretability analysis of a Transformer-based Sign Language Translation (SLT) model, focusing on the translation from video-based Greek Sign Language to glosses and text. Leveraging the Greek Sign Language Dataset, we examine the attention mechanisms within the model to understand how it processes and aligns visual input with sequential glosses. Our analysis reveals that the model pays attention to clusters of frames rather than individual ones, with a diagonal alignment pattern emerging between poses and glosses, which becomes less distinct as the number of glosses increases. We also explore the relative contributions of cross-attention and self-attention at each decoding step, finding that the model initially relies on video frames but shifts its focus to previously predicted tokens as the translation progresses. This work contributes to a deeper understanding of SLT models, paving the way for the development of more transparent and reliable translation systems essential for real-world applications.

Nicolás Gaggion, Rodrigo Echeveste, Lucas Mansilla et al.

It has recently been shown that deep learning models for anatomical segmentation in medical images can exhibit biases against certain sub-populations defined in terms of protected attributes like sex or ethnicity. In this context, auditing fairness of deep segmentation models becomes crucial. However, such audit process generally requires access to ground-truth segmentation masks for the target population, which may not always be available, especially when going from development to deployment. Here we propose a new method to anticipate model biases in biomedical image segmentation in the absence of ground-truth annotations. Our unsupervised bias discovery method leverages the reverse classification accuracy framework to estimate segmentation quality. Through numerical experiments in synthetic and realistic scenarios we show how our method is able to successfully anticipate fairness issues in the absence of ground-truth labels, constituting a novel and valuable tool in this field.

María Agustina Ricci Lara, Candelaria Mosquera, Enzo Ferrante et al.

In recent years the development of artificial intelligence (AI) systems for automated medical image analysis has gained enormous momentum. At the same time, a large body of work has shown that AI systems can systematically and unfairly discriminate against certain populations in various application scenarios. These two facts have motivated the emergence of algorithmic fairness studies in this field. Most research on healthcare algorithmic fairness to date has focused on the assessment of biases in terms of classical discrimination metrics such as AUC and accuracy. Potential biases in terms of model calibration, however, have only recently begun to be evaluated. This is especially important when working with clinical decision support systems, as predictive uncertainty is key for health professionals to optimally evaluate and combine multiple sources of information. In this work we study discrimination and calibration biases in models trained for automatic detection of malignant dermatological conditions from skin lesions images. Importantly, we show how several typically employed calibration metrics are systematically biased with respect to sample sizes, and how this can lead to erroneous fairness analysis if not taken into consideration. This is of particular relevance to fairness studies, where data imbalance results in drastic sample size differences between demographic sub-groups, which, if not taken into account, can act as confounders.

Eike Petersen, Enzo Ferrante, Melanie Ganz et al.

Medical imaging models have been shown to encode information about patient demographics such as age, race, and sex in their latent representation, raising concerns about their potential for discrimination. Here, we ask whether requiring models not to encode demographic attributes is desirable. We point out that marginal and class-conditional representation invariance imply the standard group fairness notions of demographic parity and equalized odds, respectively. In addition, however, they require matching the risk distributions, thus potentially equalizing away important group differences. Enforcing the traditional fairness notions directly instead does not entail these strong constraints. Moreover, representationally invariant models may still take demographic attributes into account for deriving predictions, implying unequal treatment - in fact, achieving representation invariance may require doing so. In theory, this can be prevented using counterfactual notions of (individual) fairness or invariance. We caution, however, that properly defining medical image counterfactuals with respect to demographic attributes is fraught with challenges. Finally, we posit that encoding demographic attributes may even be advantageous if it enables learning a task-specific encoding of demographic features that does not rely on social constructs such as 'race' and 'gender.' We conclude that demographically invariant representations are neither necessary nor sufficient for fairness in medical imaging. Models may need to encode demographic attributes, lending further urgency to calls for comprehensive model fairness assessments in terms of predictive performance across diverse patient groups.

Sean I. Young, Adrian V. Dalca, Enzo Ferrante et al.

Learning-based image reconstruction models, such as those based on the U-Net, require a large set of labeled images if good generalization is to be guaranteed. In some imaging domains, however, labeled data with pixel- or voxel-level label accuracy are scarce due to the cost of acquiring them. This problem is exacerbated further in domains like medical imaging, where there is no single ground truth label, resulting in large amounts of repeat variability in the labels. Therefore, training reconstruction networks to generalize better by learning from both labeled and unlabeled examples (called semi-supervised learning) is problem of practical and theoretical interest. However, traditional semi-supervised learning methods for image reconstruction often necessitate handcrafting a differentiable regularizer specific to some given imaging problem, which can be extremely time-consuming. In this work, we propose "supervision by denoising" (SUD), a framework that enables us to supervise reconstruction models using their own denoised output as soft labels. SUD unifies stochastic averaging and spatial denoising techniques under a spatio-temporal denoising framework and alternates denoising and model weight update steps in an optimization framework for semi-supervision. As example applications, we apply SUD to two problems arising from biomedical imaging -- anatomical brain reconstruction (3D) and cortical parcellation (2D) -- to demonstrate a significant improvement in the image reconstructions over supervised-only and stochastic averaging baselines.

Candelaria Mosquera, Luciana Ferrer, Diego Milone et al.

This work aims to analyze standard evaluation practices adopted by the research community when assessing chest x-ray classifiers, particularly focusing on the impact of class imbalance in such appraisals. Our analysis considers a comprehensive definition of model performance, covering not only discriminative performance but also model calibration, a topic of research that has received increasing attention during the last years within the machine learning community. Firstly, we conducted a literature study to analyze common scientific practices and confirmed that: (1) even when dealing with highly imbalanced datasets, the community tends to use metrics that are dominated by the majority class; and (2) it is still uncommon to include calibration studies for chest x-ray classifiers, albeit its importance in the context of healthcare. Secondly, we perform a systematic experiment on two major chest x-ray datasets to explore the behavior of several performance metrics under different class ratios and show that widely adopted metrics can conceal the performance in the minority class. Finally, we recommend the inclusion of complementary metrics to better reflect the system's performance in such scenarios. Our study indicates that current evaluation practices adopted by the research community for chest x-ray computer-aided diagnosis systems may not reflect their performance in real clinical scenarios, and suggest alternatives to improve this situation.

Agostina Larrazabal, Cesar Martinez, Jose Dolz et al.

Modern deep neural networks achieved remarkable progress in medical image segmentation tasks. However, it has recently been observed that they tend to produce overconfident estimates, even in situations of high uncertainty, leading to poorly calibrated and unreliable models. In this work we introduce Maximum Entropy on Erroneous Predictions (MEEP), a training strategy for segmentation networks which selectively penalizes overconfident predictions, focusing only on misclassified pixels. Our method is agnostic to the neural architecture, does not increase model complexity and can be coupled with multiple segmentation loss functions. We benchmark the proposed strategy in two challenging segmentation tasks: white matter hyperintensity lesions in magnetic resonance images (MRI) of the brain, and atrial segmentation in cardiac MRI. The experimental results demonstrate that coupling MEEP with standard segmentation losses leads to improvements not only in terms of model calibration, but also in segmentation quality.

Lucas Mansilla, Rodrigo Echeveste, Diego H. Milone et al.

In real-life applications, machine learning models often face scenarios where there is a change in data distribution between training and test domains. When the aim is to make predictions on distributions different from those seen at training, we incur in a domain generalization problem. Methods to address this issue learn a model using data from multiple source domains, and then apply this model to the unseen target domain. Our hypothesis is that when training with multiple domains, conflicting gradients within each mini-batch contain information specific to the individual domains which is irrelevant to the others, including the test domain. If left untouched, such disagreement may degrade generalization performance. In this work, we characterize the conflicting gradients emerging in domain shift scenarios and devise novel gradient agreement strategies based on gradient surgery to alleviate their effect. We validate our approach in image classification tasks with three multi-domain datasets, showing the value of the proposed agreement strategy in enhancing the generalization capability of deep learning models in domain shift scenarios.

Nicolás Gaggion, Lucas Mansilla, Diego Milone et al.

In this work we address the problem of landmark-based segmentation for anatomical structures. We propose HybridGNet, an encoder-decoder neural architecture which combines standard convolutions for image feature encoding, with graph convolutional neural networks to decode plausible representations of anatomical structures. We benchmark the proposed architecture considering other standard landmark and pixel-based models for anatomical segmentation in chest x-ray images, and found that HybridGNet is more robust to image occlusions. We also show that it can be used to construct landmark-based segmentations from pixel level annotations. Our experimental results suggest that HybridGNet produces accurate and anatomically plausible landmark-based segmentations, by naturally incorporating shape constraints within the decoding process via spectral convolutions.

Agostina J. Larrazabal, César Martínez, Jose Dolz et al.

Despite the astonishing performance of deep-learning based approaches for visual tasks such as semantic segmentation, they are known to produce miscalibrated predictions, which could be harmful for critical decision-making processes. Ensemble learning has shown to not only boost the performance of individual models but also reduce their miscalibration by averaging independent predictions. In this scenario, model diversity has become a key factor, which facilitates individual models converging to different functional solutions. In this work, we introduce Orthogonal Ensemble Networks (OEN), a novel framework to explicitly enforce model diversity by means of orthogonal constraints. The proposed method is based on the hypothesis that inducing orthogonality among the constituents of the ensemble will increase the overall model diversity. We resort to a new pairwise orthogonality constraint which can be used to regularize a sequential ensemble training process, resulting on improved predictive performance and better calibrated model outputs. We benchmark the proposed framework in two challenging brain lesion segmentation tasks --brain tumor and white matter hyper-intensity segmentation in MR images. The experimental results show that our approach produces more robust and well-calibrated ensemble models and can deal with challenging tasks in the context of biomedical image segmentation.

Franco Matzkin, Virginia Newcombe, Ben Glocker et al.

Cranial implant design is a challenging task, whose accuracy is crucial in the context of cranioplasty procedures. This task is usually performed manually by experts using computer-assisted design software. In this work, we propose and evaluate alternative automatic deep learning models for cranial implant reconstruction from CT images. The models are trained and evaluated using the database released by the AutoImplant challenge, and compared to a baseline implemented by the organizers. We employ a simulated virtual craniectomy to train our models using complete skulls, and compare two different approaches trained with this procedure. The first one is a direct estimation method based on the UNet architecture. The second method incorporates shape priors to increase the robustness when dealing with out-of-distribution implant shapes. Our direct estimation method outperforms the baselines provided by the organizers, while the model with shape priors shows superior performance when dealing with out-of-distribution cases. Overall, our methods show promising results in the difficult task of cranial implant design.

Julian Alberto Palladino, Diego Fernandez Slezak, Enzo Ferrante

Automatic segmentation of white matter hyperintensities in magnetic resonance images is of paramount clinical and research importance. Quantification of these lesions serve as a predictor for risk of stroke, dementia and mortality. During the last years, convolutional neural networks (CNN) specifically tailored for biomedical image segmentation have outperformed all previous techniques in this task. However, they are extremely data-dependent, and maintain a good performance only when data distribution between training and test datasets remains unchanged. When such distribution changes but we still aim at performing the same task, we incur in a domain adaptation problem (e.g. using a different MR machine or different acquisition parameters for training and test data). In this work, we explore the use of cycle-consistent adversarial networks (CycleGAN) to perform unsupervised domain adaptation on multicenter MR images with brain lesions. We aim at learning a mapping function to transform volumetric MR images between domains, which are characterized by different medical centers and MR machines with varying brand, model and configuration parameters. Our experiments show that CycleGAN allows us to reduce the Jensen-Shannon divergence between MR domains, enabling automatic segmentation with CNN models on domains where no labeled data was available.

Franco Matzkin, Virginia Newcombe, Susan Stevenson et al.

Decompressive craniectomy (DC) is a common surgical procedure consisting of the removal of a portion of the skull that is performed after incidents such as stroke, traumatic brain injury (TBI) or other events that could result in acute subdural hemorrhage and/or increasing intracranial pressure. In these cases, CT scans are obtained to diagnose and assess injuries, or guide a certain therapy and intervention. We propose a deep learning based method to reconstruct the skull defect removed during DC performed after TBI from post-operative CT images. This reconstruction is useful in multiple scenarios, e.g. to support the creation of cranioplasty plates, accurate measurements of bone flap volume and total intracranial volume, important for studies that aim to relate later atrophy to patient outcome. We propose and compare alternative self-supervised methods where an encoder-decoder convolutional neural network (CNN) estimates the missing bone flap on post-operative CTs. The self-supervised learning strategy only requires images with complete skulls and avoids the need for annotated DC images. For evaluation, we employ real and simulated images with DC, comparing the results with other state-of-the-art approaches. The experiments show that the proposed model outperforms current manual methods, enabling reconstruction even in highly challenging cases where big skull defects have been removed during surgery.

Agostina J Larrazabal, César Martínez, Ben Glocker et al.

We introduce Post-DAE, a post-processing method based on denoising autoencoders (DAE) to improve the anatomical plausibility of arbitrary biomedical image segmentation algorithms. Some of the most popular segmentation methods (e.g. based on convolutional neural networks or random forest classifiers) incorporate additional post-processing steps to ensure that the resulting masks fulfill expected connectivity constraints. These methods operate under the hypothesis that contiguous pixels with similar aspect should belong to the same class. Even if valid in general, this assumption does not consider more complex priors like topological restrictions or convexity, which cannot be easily incorporated into these methods. Post-DAE leverages the latest developments in manifold learning via denoising autoencoders. First, we learn a compact and non-linear embedding that represents the space of anatomically plausible segmentations. Then, given a segmentation mask obtained with an arbitrary method, we reconstruct its anatomically plausible version by projecting it onto the learnt manifold. The proposed method is trained using unpaired segmentation mask, what makes it independent of intensity information and image modality. We performed experiments in binary and multi-label segmentation of chest X-ray and cardiac magnetic resonance images. We show how erroneous and noisy segmentation masks can be improved using Post-DAE. With almost no additional computation cost, our method brings erroneous segmentations back to a feasible space.

Lucas Mansilla, Diego H. Milone, Enzo Ferrante

Deformable image registration is a fundamental problem in the field of medical image analysis. During the last years, we have witnessed the advent of deep learning-based image registration methods which achieve state-of-the-art performance, and drastically reduce the required computational time. However, little work has been done regarding how can we encourage our models to produce not only accurate, but also anatomically plausible results, which is still an open question in the field. In this work, we argue that incorporating anatomical priors in the form of global constraints into the learning process of these models, will further improve their performance and boost the realism of the warped images after registration. We learn global non-linear representations of image anatomy using segmentation masks, and employ them to constraint the registration process. The proposed AC-RegNet architecture is evaluated in the context of chest X-ray image registration using three different datasets, where the high anatomical variability makes the task extremely challenging. Our experiments show that the proposed anatomically constrained registration model produces more realistic and accurate results than state-of-the-art methods, demonstrating the potential of this approach.

Agostina J. Larrazabal, Cesar Martinez, Enzo Ferrante

Deep convolutional neural networks (CNN) proved to be highly accurate to perform anatomical segmentation of medical images. However, some of the most popular CNN architectures for image segmentation still rely on post-processing strategies (e.g. Conditional Random Fields) to incorporate connectivity constraints into the resulting masks. These post-processing steps are based on the assumption that objects are usually continuous and therefore nearby pixels should be assigned the same object label. Even if it is a valid assumption in general, these methods do not offer a straightforward way to incorporate more complex priors like convexity or arbitrary shape restrictions. In this work we propose Post-DAE, a post-processing method based on denoising autoencoders (DAE) trained using only segmentation masks. We learn a low-dimensional space of anatomically plausible segmentations, and use it as a post-processing step to impose shape constraints on the resulting masks obtained with arbitrary segmentation methods. Our approach is independent of image modality and intensity information since it employs only segmentation masks for training. This enables the use of anatomical segmentations that do not need to be paired with intensity images, making the approach very flexible. Our experimental results on anatomical segmentation of X-ray images show that Post-DAE can improve the quality of noisy and incorrect segmentation masks obtained with a variety of standard methods, by bringing them back to a feasible space, with almost no extra computational time.

Nicolas Roulet, Diego Fernandez Slezak, Enzo Ferrante

Brain lesion and anatomy segmentation in magnetic resonance images are fundamental tasks in neuroimaging research and clinical practice. Given enough training data, convolutional neuronal networks (CNN) proved to outperform all existent techniques in both tasks independently. However, to date, little work has been done regarding simultaneous learning of brain lesion and anatomy segmentation from disjoint datasets. In this work we focus on training a single CNN model to predict brain tissue and lesion segmentations using heterogeneous datasets labeled independently, according to only one of these tasks (a common scenario when using publicly available datasets). We show that label contradiction issues can arise in this case, and propose a novel adaptive cross entropy (ACE) loss function that makes such training possible. We provide quantitative evaluation in two different scenarios, benchmarking the proposed method in comparison with a multi-network approach. Our experiments suggest that ACE loss enables training of single models when standard cross entropy and Dice loss functions tend to fail. Moreover, we show that it is possible to achieve competitive results when comparing with multiple networks trained for independent tasks.

Spyridon Bakas, Mauricio Reyes, Andras Jakab et al.

Gliomas are the most common primary brain malignancies, with different degrees of aggressiveness, variable prognosis and various heterogeneous histologic sub-regions, i.e., peritumoral edematous/invaded tissue, necrotic core, active and non-enhancing core. This intrinsic heterogeneity is also portrayed in their radio-phenotype, as their sub-regions are depicted by varying intensity profiles disseminated across multi-parametric magnetic resonance imaging (mpMRI) scans, reflecting varying biological properties. Their heterogeneous shape, extent, and location are some of the factors that make these tumors difficult to resect, and in some cases inoperable. The amount of resected tumor is a factor also considered in longitudinal scans, when evaluating the apparent tumor for potential diagnosis of progression. Furthermore, there is mounting evidence that accurate segmentation of the various tumor sub-regions can offer the basis for quantitative image analysis towards prediction of patient overall survival. This study assesses the state-of-the-art machine learning (ML) methods used for brain tumor image analysis in mpMRI scans, during the last seven instances of the International Brain Tumor Segmentation (BraTS) challenge, i.e., 2012-2018. Specifically, we focus on i) evaluating segmentations of the various glioma sub-regions in pre-operative mpMRI scans, ii) assessing potential tumor progression by virtue of longitudinal growth of tumor sub-regions, beyond use of the RECIST/RANO criteria, and iii) predicting the overall survival from pre-operative mpMRI scans of patients that underwent gross total resection. Finally, we investigate the challenge of identifying the best ML algorithms for each of these tasks, considering that apart from being diverse on each instance of the challenge, the multi-institutional mpMRI BraTS dataset has also been a continuously evolving/growing dataset.