Aasa Feragen

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.

Eike Petersen, Aasa Feragen, Maria Luise da Costa Zemsch et al.

Convolutional neural networks have enabled significant improvements in medical image-based diagnosis. It is, however, increasingly clear that these models are susceptible to performance degradation when facing spurious correlations and dataset shift, leading, e.g., to underperformance on underrepresented patient groups. In this paper, we compare two classification schemes on the ADNI MRI dataset: a simple logistic regression model using manually selected volumetric features, and a convolutional neural network trained on 3D MRI data. We assess the robustness of the trained models in the face of varying dataset splits, training set sex composition, and stage of disease. In contrast to earlier work in other imaging modalities, we do not observe a clear pattern of improved model performance for the majority group in the training dataset. Instead, while logistic regression is fully robust to dataset composition, we find that CNN performance is generally improved for both male and female subjects when including more female subjects in the training dataset. We hypothesize that this might be due to inherent differences in the pathology of the two sexes. Moreover, in our analysis, the logistic regression model outperforms the 3D CNN, emphasizing the utility of manual feature specification based on prior knowledge, and the need for more robust automatic feature selection.

Eike Petersen, Melanie Ganz, Sune Hannibal Holm et al.

Recent work on algorithmic fairness has largely focused on the fairness of discrete decisions, or classifications. While such decisions are often based on risk score models, the fairness of the risk models themselves has received considerably less attention. Risk models are of interest for a number of reasons, including the fact that they communicate uncertainty about the potential outcomes to users, thus representing a way to enable meaningful human oversight. Here, we address fairness desiderata for risk score models. We identify the provision of similar epistemic value to different groups as a key desideratum for risk score fairness. Further, we address how to assess the fairness of risk score models quantitatively, including a discussion of metric choices and meaningful statistical comparisons between groups. In this context, we also introduce a novel calibration error metric that is less sample size-biased than previously proposed metrics, enabling meaningful comparisons between groups of different sizes. We illustrate our methodology - which is widely applicable in many other settings - in two case studies, one in recidivism risk prediction, and one in risk of major depressive disorder (MDD) prediction.

Manxi Lin, Zahra Bashir, Martin Grønnebæk Tolsgaard et al.

Curvilinear structure segmentation is important in medical imaging, quantifying structures such as vessels, airways, neurons, or organ boundaries in 2D slices. Segmentation via pixel-wise classification often fails to capture the small and low-contrast curvilinear structures. Prior topological information is typically used to address this problem, often at an expensive computational cost, and sometimes requiring prior knowledge of the expected topology. We present DTU-Net, a data-driven approach to topology-preserving curvilinear structure segmentation. DTU-Net consists of two sequential, lightweight U-Nets, dedicated to texture and topology, respectively. While the texture net makes a coarse prediction using image texture information, the topology net learns topological information from the coarse prediction by employing a triplet loss trained to recognize false and missed splits in the structure. We conduct experiments on a challenging multi-class ultrasound scan segmentation dataset as well as a well-known retinal imaging dataset. Results show that our model outperforms existing approaches in both pixel-wise segmentation accuracy and topological continuity, with no need for prior topological knowledge.

Nina Weng, Siavash Bigdeli, Eike Petersen et al.

While many studies have assessed the fairness of AI algorithms in the medical field, the causes of differences in prediction performance are often unknown. This lack of knowledge about the causes of bias hampers the efficacy of bias mitigation, as evidenced by the fact that simple dataset balancing still often performs best in reducing performance gaps but is unable to resolve all performance differences. In this work, we investigate the causes of gender bias in machine learning-based chest X-ray diagnosis. In particular, we explore the hypothesis that breast tissue leads to underexposure of the lungs and causes lower model performance. Methodologically, we propose a new sampling method which addresses the highly skewed distribution of recordings per patient in two widely used public datasets, while at the same time reducing the impact of label errors. Our comprehensive analysis of gender differences across diseases, datasets, and gender representations in the training set shows that dataset imbalance is not the sole cause of performance differences. Moreover, relative group performance differs strongly between datasets, indicating important dataset-specific factors influencing male/female group performance. Finally, we investigate the effect of breast tissue more specifically, by cropping out the breasts from recordings, finding that this does not resolve the observed performance gaps. In conclusion, our results indicate that dataset-specific factors, not fundamental physiological differences, are the main drivers of male--female performance gaps in chest X-ray analyses on widely used NIH and CheXpert Dataset.

Kilian Zepf, Eike Petersen, Jes Frellsen et al.

Segmentation uncertainty models predict a distribution over plausible segmentations for a given input, which they learn from the annotator variation in the training set. However, in practice these annotations can differ systematically in the way they are generated, for example through the use of different labeling tools. This results in datasets that contain both data variability and differing label styles. In this paper, we demonstrate that applying state-of-the-art segmentation uncertainty models on such datasets can lead to model bias caused by the different label styles. We present an updated modelling objective conditioning on labeling style for aleatoric uncertainty estimation, and modify two state-of-the-art-architectures for segmentation uncertainty accordingly. We show with extensive experiments that this method reduces label style bias, while improving segmentation performance, increasing the applicability of segmentation uncertainty models in the wild. We curate two datasets, with annotations in different label styles, which we will make publicly available along with our code upon publication.

Kamil Mikolaj, Manxi Lin, Zahra Bashir et al.

Confounding information in the form of text or markings embedded in medical images can severely affect the training of diagnostic deep learning algorithms. However, data collected for clinical purposes often have such markings embedded in them. In dermatology, known examples include drawings or rulers that are overrepresented in images of malignant lesions. In this paper, we encounter text and calipers placed on the images found in national databases containing fetal screening ultrasound scans, which correlate with standard planes to be predicted. In order to utilize the vast amounts of data available in these databases, we develop and validate a series of methods for minimizing the confounding effects of embedded text and calipers on deep learning algorithms designed for ultrasound, using standard plane classification as a test case.

Kilian Zepf, Selma Wanna, Marco Miani et al.

Image segmentation relies heavily on neural networks which are known to be overconfident, especially when making predictions on out-of-distribution (OOD) images. This is a common scenario in the medical domain due to variations in equipment, acquisition sites, or image corruptions. This work addresses the challenge of OOD detection by proposing Laplacian Segmentation Networks (LSN): methods which jointly model epistemic (model) and aleatoric (data) uncertainty for OOD detection. In doing so, we propose the first Laplace approximation of the weight posterior that scales to large neural networks with skip connections that have high-dimensional outputs. We demonstrate on three datasets that the LSN-modeled parameter distributions, in combination with suitable uncertainty measures, gives superior OOD detection.

Manxi Lin, Aasa Feragen, Kamil Mikolaj et al.

The quality of fetal ultrasound screening scans directly influences the precision of biometric measurements. However, acquiring high-quality scans is labor-intensive and highly relies on the operator's skills. Considering the low contrastiveness and imaging artifacts that widely exist in ultrasound, even a dedicated deep-learning model can be vulnerable to learning from confounding information in the image. In this paper, we propose a holistic and explainable method for fetal ultrasound quality assessment, where we design a hierarchical concept bottleneck model by introducing human-readable ``concepts" into the task and imitating the sequential expert decision-making process. This hierarchical information flow forces the model to learn concepts from semantically meaningful areas: The model first passes through a layer of visual, segmentation-based concepts, and next a second layer of property concepts directly associated with the decision-making task. We consider the quality assessment to be in a more challenging but more realistic setting, with fine-grained image recognition. Experiments show that our model outperforms equivalent concept-free models on an in-house dataset, and shows better generalizability on two public benchmarks, one from Spain and one from Africa, without any fine-tuning.

Kilian Zepf, Jes Frellsen, Aasa Feragen

We address the selection and evaluation of uncertain segmentation methods in medical imaging and present two case studies: prostate segmentation, illustrating that for minimal annotator variation simple deterministic models can suffice, and lung lesion segmentation, highlighting the limitations of the Generalized Energy Distance (GED) in model selection. Our findings lead to guidelines for accurately choosing and developing uncertain segmentation models, that integrate aleatoric and epistemic components. These guidelines are designed to aid researchers and practitioners in better developing, selecting, and evaluating uncertain segmentation methods, thereby facilitating enhanced adoption and effective application of segmentation uncertainty in practice.

Chun Kit Wong, Manxi Lin, Alberto Raheli et al.

Examination of the umbilical artery with Doppler ultrasonography is performed to investigate blood supply to the fetus through the umbilical cord, which is vital for the monitoring of fetal health. Such examination involves several steps that must be performed correctly: identifying suitable sites on the umbilical artery for the measurement, acquiring the blood flow curve in the form of a Doppler spectrum, and ensuring compliance to a set of quality standards. These steps rely heavily on the operator's skill, and the shortage of experienced sonographers has thus created a demand for machine assistance. In this work, we propose an automatic system to fill the gap. By using a modified Faster R-CNN network, we obtain an algorithm that can suggest locations suitable for Doppler measurement. Meanwhile, we have also developed a method for assessment of the Doppler spectrum's quality. The proposed system is validated on 657 images from a national ultrasound screening database, with results demonstrating its potential as a guidance system.

Aditya Parikh, Sneha Das, Aasa Feragen

Algorithmic bias in medical imaging can perpetuate health disparities, yet its causes remain poorly understood in segmentation tasks. While fairness has been extensively studied in classification, segmentation remains underexplored despite its clinical importance. In breast cancer segmentation, models exhibit significant performance disparities against younger patients, commonly attributed to physiological differences in breast density. We audit the MAMA-MIA dataset, establishing a quantitative baseline of age-related bias in its automated labels, and reveal a critical Biased Ruler effect where systematically flawed labels for validation misrepresent a model's actual bias. However, whether this bias originates from lower-quality annotations (label bias) or from fundamentally more challenging image characteristics remains unclear. Through controlled experiments, we systematically refute hypotheses that the bias stems from label quality sensitivity or quantitative case difficulty imbalance. Balancing training data by difficulty fails to mitigate the disparity, revealing that younger patient cases are intrinsically harder to learn. We provide direct evidence that systemic bias is learned and amplified when training on biased, machine-generated labels, a critical finding for automated annotation pipelines. This work introduces a systematic framework for diagnosing algorithmic bias in medical segmentation and demonstrates that achieving fairness requires addressing qualitative distributional differences rather than merely balancing case counts.

Aditya Parikh, Sneha Das, Aasa Feragen

Deep learning models aim to improve diagnostic workflows, but fairness evaluation remains underexplored beyond classification, e.g., in image segmentation. Unaddressed segmentation bias can lead to disparities in the quality of care for certain populations, potentially compounded across clinical decision points and amplified through iterative model development. Here, we audit the fairness of the automated segmentation labels provided in the breast cancer tumor segmentation dataset MAMA-MIA. We evaluate automated segmentation quality across age, ethnicity, and data source. Our analysis reveals an intrinsic age-related bias against younger patients that continues to persist even after controlling for confounding factors, such as data source. We hypothesize that this bias may be linked to physiological factors, a known challenge for both radiologists and automated systems. Finally, we show how aggregating data from multiple data sources influences site-specific ethnic biases, underscoring the necessity of investigating data at a granular level.

Nina Weng, Paraskevas Pegios, Eike Petersen et al.

Shortcut learning is when a model -- e.g. a cardiac disease classifier -- exploits correlations between the target label and a spurious shortcut feature, e.g. a pacemaker, to predict the target label based on the shortcut rather than real discriminative features. This is common in medical imaging, where treatment and clinical annotations correlate with disease labels, making them easy shortcuts to predict disease. We propose a novel detection and quantification of the impact of potential shortcut features via a fast diffusion-based counterfactual image generation that can synthetically remove or add shortcuts. Via a novel inpainting-based modification we spatially limit the changes made with no extra inference step, encouraging the removal of spatially constrained shortcut features while ensuring that the shortcut-free counterfactuals preserve their remaining image features to a high degree. Using these, we assess how shortcut features influence model predictions. This is enabled by our second contribution: An efficient diffusion-based counterfactual explanation method with significant inference speed-up at comparable image quality as state-of-the-art. We confirm this on two large chest X-ray datasets, a skin lesion dataset, and CelebA. Our code is publicly available at fastdime.compute.dtu.dk.

Manxi Lin, Nina Weng, Kamil Mikolaj et al.

Shortcut learning is a phenomenon where machine learning models prioritize learning simple, potentially misleading cues from data that do not generalize well beyond the training set. While existing research primarily investigates this in the realm of image classification, this study extends the exploration of shortcut learning into medical image segmentation. We demonstrate that clinical annotations such as calipers, and the combination of zero-padded convolutions and center-cropped training sets in the dataset can inadvertently serve as shortcuts, impacting segmentation accuracy. We identify and evaluate the shortcut learning on two different but common medical image segmentation tasks. In addition, we suggest strategies to mitigate the influence of shortcut learning and improve the generalizability of the segmentation models. By uncovering the presence and implications of shortcuts in medical image segmentation, we provide insights and methodologies for evaluating and overcoming this pervasive challenge and call for attention in the community for shortcuts in segmentation. Our code is public at https://github.com/nina-weng/shortcut_skinseg .

Tiantian Zhang, Manxi Lin, Hongda Guo et al.

The Prostate Imaging Reporting and Data System (PI-RADS) is pivotal in the diagnosis of clinically significant prostate cancer through MRI imaging. Current deep learning-based PI-RADS scoring methods often lack the incorporation of common PI-RADS clinical guideline~(PICG) utilized by radiologists, potentially compromising scoring accuracy. This paper introduces a novel approach that adapts a multi-modal large language model (MLLM) to incorporate PICG into PI-RADS scoring model without additional annotations and network parameters. We present a designed two-stage fine-tuning process aiming at adapting a MLLM originally trained on natural images to the MRI images while effectively integrating the PICG. Specifically, in the first stage, we develop a domain adapter layer tailored for processing 3D MRI inputs and instruct the MLLM to differentiate MRI sequences. In the second stage, we translate PICG for guiding instructions from the model to generate PICG-guided image features. Through such a feature distillation step, we align the scoring network's features with the PICG-guided image features, which enables the model to effectively incorporate the PICG information. We develop our model on a public dataset and evaluate it on an in-house dataset. Experimental results demonstrate that our approach effectively improves the performance of current scoring networks. Code is available at: https://github.com/med-air/PICG2scoring

Markus Ditlev Sjøgren Olsen, Jakob Ambsdorf, Manxi Lin et al.

Congenital malformations of the brain are among the most common fetal abnormalities that impact fetal development. Previous anomaly detection methods on ultrasound images are based on supervised learning, rely on manual annotations, and risk missing underrepresented categories. In this work, we frame fetal brain anomaly detection as an unsupervised task using diffusion models. To this end, we employ an inpainting-based Noise Agnostic Anomaly Detection approach that identifies the abnormality using diffusion-reconstructed fetal brain images from multiple noise levels. Our approach only requires normal fetal brain ultrasound images for training, addressing the limited availability of abnormal data. Our experiments on a real-world clinical dataset show the potential of using unsupervised methods for fetal brain anomaly detection. Additionally, we comprehensively evaluate how different noise types affect diffusion models in the fetal anomaly detection domain.

Aditya Parikh, Aasa Feragen, Sneha Das et al.

Reliable deployment of Vision-Language Models (VLMs) in radiology requires validation metrics that go beyond surface-level text similarity to ensure clinical fidelity and demographic fairness. This paper investigates a critical blind spot in current model evaluation: the use of decoding strategies that lead to high aggregate token-overlap scores despite succumbing to template collapse, in which models generate only repetitive, safe generic text and omit clinical terminology. Unaddressed, this blind spot can lead to metric gaming, where models that perform well on benchmarks prove clinically uninformative. Instead, we advocate for lexical diversity measures to check model generations for clinical specificity. We introduce Clinical Association Displacement (CAD), a vocabulary-level framework that quantifies shifts in demographic-based word associations in generated reports. Weighted Association Erasure (WAE) aggregates these shifts to measure the clinical signal loss across demographic groups. We show that deterministic decoding produces high levels of semantic erasure, while stochastic sampling generates diverse outputs but risks introducing new bias, motivating a fundamental rethink of how "optimal" reporting is defined.

Aditya Parikh, Aasa Feragen

We present a fairness-aware framework for multi-class lung disease diagnosis from chest CT volumes, developed for the Fair Disease Diagnosis Challenge at the PHAROS-AIF-MIH Workshop (CVPR 2026). The challenge requires classifying CT scans into four categories -- Healthy, COVID-19, Adenocarcinoma, and Squamous Cell Carcinoma -- with performance measured as the average of per-gender macro F1 scores, explicitly penalizing gender-inequitable predictions. Our approach addresses two core difficulties: the sparse pathological signal across hundreds of slices, and a severe demographic imbalance compounded across disease class and gender. We propose an attention-based Multiple Instance Learning (MIL) model on a ConvNeXt backbone that learns to identify diagnostically relevant slices without slice-level supervision, augmented with a Gradient Reversal Layer (GRL) that adversarially suppresses gender-predictive structure in the learned scan representation. Training incorporates focal loss with label smoothing, stratified cross-validation over joint (class, gender) strata, and targeted oversampling of the most underrepresented subgroup. At inference, all five-fold checkpoints are ensembled with horizontal-flip test-time augmentation via soft logit voting and out-of-the-fold threshold optimization for robustness. Our model achieves a mean validation competition score of 0.685 (std - 0.030), with the best single fold reaching 0.759. All training and inference code is publicly available at https://github.com/ADE-17/cvpr-fair-chest-ct

Nina Weng, Aasa Feragen, Siavash Bigdeli

Diffusion-based generative models, such as Denoising Diffusion Probabilistic Models (DDPMs), have achieved remarkable success in image generation, but their step-by-step denoising process remains opaque, leaving critical aspects of the generation mechanism unexplained. To address this, we introduce \emph{Patronus}, an interpretable diffusion model inspired by ProtoPNet. Patronus integrates a prototypical network into DDPMs, enabling the extraction of prototypes and conditioning of the generation process on their prototype activation vector. This design enhances interpretability by showing the learned prototypes and how they influence the generation process. Additionally, the model supports downstream tasks like image manipulation, enabling more transparent and controlled modifications. Moreover, Patronus could reveal shortcut learning in the generation process by detecting unwanted correlations between learned prototypes. Notably, Patronus operates entirely without any annotations or text prompts. This work opens new avenues for understanding and controlling diffusion models through prototype-based interpretability. Our code is available at \href{https://github.com/nina-weng/patronus}{https://github.com/nina-weng/patronus}.

Aditya Parikh, Stella Frank, Sneha Das et al.

Labeled datasets reflect the biases of their annotation pipelines, which sometimes introduce label bias: group-conditional label errors that cause systematic performance disparities across demographic subgroups. Label bias in image segmentation remains underexplored, as even detecting it typically requires clean, unbiased annotations, which are not readily available. We present a data-centric adaptation of Confident Learning to segmentation, allowing detection of label bias directly in the training data without a clean, unbiased ground truth. By comparing the provided training labels to the model's confident predictions, we isolate directional errors that quantify the presence and nature of bias, where standard overlap metrics like Dice fail. We further show that label bias influences subgroup separability in the encoder's feature space, an artifact we leverage for bias mitigation rather than suppressing it. We evaluate three datasets, spanning from synthetic to real-life bias, showing how our framework reliably detects and mitigates bias without access to clean labels, achieving equitable performance across experimental conditions.

Aya Elgebaly, Joris Fournel, Benjamin Laine Jønch Jurgensen et al.

Bias in medical AI is often framed as a problem of representation. However, in image-based tasks such as fetal ultrasound, performance disparities can arise even when representation is adequate, because predictive accuracy depends strongly on image quality. Image quality is shaped by acquisition conditions and operator expertise, as well as patient-dependent factors such as maternal body mass index (BMI), all of which may correlate with sensitive demographic features. Consequently, observed disparities may reflect the combined influence of demographic, clinical, and acquisition-related factors rather than data imbalance alone, and may obscure underlying interaction or confounding effects. We propose a structured framework to explore and detect intersectional bias, combining unsupervised slice discovery, systematic factor-wise analysis, and targeted intersectional evaluation. In a case study of over 94{,}000 ultrasound images for fetal weight estimation, we analyze bias in a state-of-the-art deep learning (DL) model and the clinical standard Hadlock, a regression formula using biometric measurements. Pixel spacing (PS) -- a parameter considered suboptimal in current acquisition protocols -- emerged as a consistent driver of performance differences, with higher PS associated with improvements of up to 24\% in selected subgroups for both models. Because PS is often adapted in cases of high BMI or low gestational age (GA), this effect carries a substantial risk of confounding. Our intersectional analysis revealed that part of the PS-associated signal is explained by GA, while PS-related improvements persist across BMI strata, highlighting the importance of acquisition-aware and interaction-aware evaluation in medical AI fairness research.

Paraskevas Pegios, Manxi Lin, Nina Weng et al.

Obstetric ultrasound image quality is crucial for accurate diagnosis and monitoring of fetal health. However, acquiring high-quality standard planes is difficult, influenced by the sonographer's expertise and factors like the maternal BMI or fetus dynamics. In this work, we explore diffusion-based counterfactual explainable AI to generate realistic, high-quality standard planes from low-quality non-standard ones. Through quantitative and qualitative evaluation, we demonstrate the effectiveness of our approach in generating plausible counterfactuals of increased quality. This shows future promise for enhancing training of clinicians by providing visual feedback and potentially improving standard plane quality and acquisition for downstream diagnosis and monitoring.

Miguel López-Pérez, Søren Hauberg, Aasa Feragen

Racial bias in medicine, such as in dermatology, presents significant ethical and clinical challenges. This is likely to happen because there is a significant underrepresentation of darker skin tones in training datasets for machine learning models. While efforts to address bias in dermatology have focused on improving dataset diversity and mitigating disparities in discriminative models, the impact of racial bias on generative models remains underexplored. Generative models, such as Variational Autoencoders (VAEs), are increasingly used in healthcare applications, yet their fairness across diverse skin tones is currently not well understood. In this study, we evaluate the fairness of generative models in clinical dermatology with respect to racial bias. For this purpose, we first train a VAE with a perceptual loss to generate and reconstruct high-quality skin images across different skin tones. We utilize the Fitzpatrick17k dataset to examine how racial bias influences the representation and performance of these models. Our findings indicate that VAE performance is, as expected, influenced by representation, i.e. increased skin tone representation comes with increased performance on the given skin tone. However, we also observe, even independently of representation, that the VAE performs better for lighter skin tones. Additionally, the uncertainty estimates produced by the VAE are ineffective in assessing the model's fairness. These results highlight the need for more representative dermatological datasets, but also a need for better understanding the sources of bias in such model, as well as improved uncertainty quantification mechanisms to detect and address racial bias in generative models for trustworthy healthcare technologies.

Paraskevas Pegios, Aasa Feragen, Andreas Abildtrup Hansen et al.

The adoption of increasingly complex deep models has fueled an urgent need for insight into how these models make predictions. Counterfactual explanations form a powerful tool for providing actionable explanations to practitioners. Previously, counterfactual explanation methods have been designed by traversing the latent space of generative models. Yet, these latent spaces are usually greatly simplified, with most of the data distribution complexity contained in the decoder rather than the latent embedding. Thus, traversing the latent space naively without taking the nonlinear decoder into account can lead to unnatural counterfactual trajectories. We introduce counterfactual explanations obtained using a Riemannian metric pulled back via the decoder and the classifier under scrutiny. This metric encodes information about the complex geometric structure of the data and the learned representation, enabling us to obtain robust counterfactual trajectories with high fidelity, as demonstrated by our experiments in real-world tabular datasets.

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/.

Manxi Lin, Jakob Ambsdorf, Emilie Pi Fogtmann Sejer et al.

We introduce the notion of semantic image quality for applications where image quality relies on semantic requirements. Working in fetal ultrasound, where ranking is challenging and annotations are noisy, we design a robust coarse-to-fine model that ranks images based on their semantic image quality and endow our predicted rankings with an uncertainty estimate. To annotate rankings on training data, we design an efficient ranking annotation scheme based on the merge sort algorithm. Finally, we compare our ranking algorithm to a number of state-of-the-art ranking algorithms on a challenging fetal ultrasound quality assessment task, showing the superior performance of our method on the majority of rank correlation metrics.

Jakob Ambsdorf, Asbjørn Munk, Sebastian Llambias et al.

With access to large-scale, unlabeled medical datasets, researchers are confronted with two questions: Should they attempt to pretrain a custom foundation model on this medical data, or use transfer-learning from an existing generalist model? And, if a custom model is pretrained, are novel methods required? In this paper we explore these questions by conducting a case-study, in which we train a foundation model on a large regional fetal ultrasound dataset of 2M images. By selecting the well-established DINOv2 method for pretraining, we achieve state-of-the-art results on three fetal ultrasound datasets, covering data from different countries, classification, segmentation, and few-shot tasks. We compare against a series of models pretrained on natural images, ultrasound images, and supervised baselines. Our results demonstrate two key insights: (i) Pretraining on custom data is worth it, even if smaller models are trained on less data, as scaling in natural image pretraining does not translate to ultrasound performance. (ii) Well-tuned methods from computer vision are making it feasible to train custom foundation models for a given medical domain, requiring no hyperparameter tuning and little methodological adaptation. Given these findings, we argue that a bias towards methodological innovation should be avoided when developing domain specific foundation models under common computational resource constraints.

Andreas Abildtrup Hansen, Paraskevas Pegios, Anna Calissano et al.

Explaining the predictions of a deep neural network is a nontrivial task, yet high-quality explanations for predictions are often a prerequisite for practitioners to trust these models. Counterfactual explanations aim to explain predictions by finding the ''nearest'' in-distribution alternative input whose prediction changes in a pre-specified way. However, it remains an open question how to define this nearest alternative input, whose solution depends on both the domain (e.g. images, graphs, tabular data, etc.) and the specific application considered. For graphs, this problem is complicated i) by their discrete nature, as opposed to the continuous nature of state-of-the-art graph classifiers; and ii) by the node permutation group acting on the graphs. We propose a method to generate counterfactual explanations for any differentiable black-box graph classifier, utilizing a case-specific permutation equivariant graph variational autoencoder. We generate counterfactual explanations in a continuous fashion by traversing the latent space of the autoencoder across the classification boundary of the classifier, allowing for seamless integration of discrete graph structure and continuous graph attributes. We empirically validate the approach on three graph datasets, showing that our model is consistently high-performing and more robust than the baselines.

Andreas Abildtrup Hansen, Anna Calissano, Aasa Feragen

Latent representations are used extensively for downstream tasks, such as visualization, interpolation or feature extraction of deep learning models. Invariant and equivariant neural networks are powerful and well-established models for enforcing inductive biases. In this paper, we demonstrate that the inductive bias imposed on the by an equivariant model must also be taken into account when using latent representations. We show how not accounting for the inductive biases leads to decreased performance on downstream tasks, and vice versa, how accounting for inductive biases can be done effectively by using an invariant projection of the latent representations. We propose principles for how to choose such a projection, and show the impact of using these principles in two common examples: First, we study a permutation equivariant variational auto-encoder trained for molecule graph generation; here we show that invariant projections can be designed that incur no loss of information in the resulting invariant representation. Next, we study a rotation-equivariant representation used for image classification. Here, we illustrate how random invariant projections can be used to obtain an invariant representation with a high degree of retained information. In both cases, the analysis of invariant latent representations proves superior to their equivariant counterparts. Finally, we illustrate that the phenomena documented here for equivariant neural networks have counterparts in standard neural networks where invariance is encouraged via augmentation. Thus, while these ambiguities may be known by experienced developers of equivariant models, we make both the knowledge as well as effective tools to handle the ambiguities available to the broader community.

Chun Kit Wong, Mary Ngo, Manxi Lin et al.

Despite the rapid development of AI models in medical image analysis, their validation in real-world clinical settings remains limited. To address this, we introduce a generic framework designed for deploying image-based AI models in such settings. Using this framework, we deployed a trained model for fetal ultrasound standard plane detection, and evaluated it in real-time sessions with both novice and expert users. Feedback from these sessions revealed that while the model offers potential benefits to medical practitioners, the need for navigational guidance was identified as a key area for improvement. These findings underscore the importance of early deployment of AI models in real-world settings, leading to insights that can guide the refinement of the model and system based on actual user feedback.

Sara Sterlie, Nina Weng, Aasa Feragen

Generative AI, such as large language models, has undergone rapid development within recent years. As these models become increasingly available to the public, concerns arise about perpetuating and amplifying harmful biases in applications. Gender stereotypes can be harmful and limiting for the individuals they target, whether they consist of misrepresentation or discrimination. Recognizing gender bias as a pervasive societal construct, this paper studies how to uncover and quantify the presence of gender biases in generative language models. In particular, we derive generative AI analogues of three well-known non-discrimination criteria from classification, namely independence, separation and sufficiency. To demonstrate these criteria in action, we design prompts for each of the criteria with a focus on occupational gender stereotype, specifically utilizing the medical test to introduce the ground truth in the generative AI context. Our results address the presence of occupational gender bias within such conversational language models.

Chun Kit Wong, Paraskevas Pegios, Nina Weng et al.

Deep learning models in medical imaging are susceptible to shortcut learning, relying on confounding metadata (e.g., scanner model) that is often encoded in image embeddings. The crucial question is whether the model actively utilizes this encoded information for its final prediction. We introduce Weight Space Correlation Analysis, an interpretable methodology that quantifies feature utilization by measuring the alignment between the classification heads of a primary clinical task and auxiliary metadata tasks. We first validate our method by successfully detecting artificially induced shortcut learning. We then apply it to probe the feature utilization of an SA-SonoNet model trained for Spontaneous Preterm Birth (sPTB) prediction. Our analysis confirmed that while the embeddings contain substantial metadata, the sPTB classifier's weight vectors were highly correlated with clinically relevant factors (e.g., birth weight) but decoupled from clinically irrelevant acquisition factors (e.g. scanner). Our methodology provides a tool to verify model trustworthiness, demonstrating that, in the absence of induced bias, the clinical model selectively utilizes features related to the genuine clinical signal.

Martina Mattioli, Eike Petersen, Aasa Feragen et al.

Artificial intelligence (AI) is being applied in almost every field. At the same time, the currently dominant deep learning methods are fundamentally black-box systems that lack explanations for their inferences, significantly limiting their trustworthiness and adoption. Explainable AI (XAI) methods aim to overcome this challenge by providing explanations of the models' decision process. Such methods are often proposed and developed by engineers and scientists with a predominantly technical background and incorporate their assumptions about the existence, validity, and explanatory utility of different conceivable explanatory mechanisms. However, the basic concept of an explanation -- what it is, whether we can know it, whether it is absolute or relative -- is far from trivial and has been the subject of deep philosophical debate for millennia. As we point out here, the assumptions incorporated into different XAI methods are not harmless and have important consequences for the validity and interpretation of AI explanations in different domains. We investigate ontological and epistemological assumptions in explainability methods when they are applied to AI systems, meaning the assumptions we make about the existence of explanations and our ability to gain knowledge about those explanations. Our analysis shows how seemingly small technical changes to an XAI method may correspond to important differences in the underlying assumptions about explanations. We furthermore highlight the risks of ignoring the underlying onto-epistemological paradigm when choosing an XAI method for a given application, and we discuss how to select and adapt appropriate XAI methods for different domains of application.

Chun Kit Wong, Anders N. Christensen, Cosmin I. Bercea et al.

Reliable out-of-distribution (OOD) detection is important for safe deployment of deep learning models in fetal ultrasound amidst heterogeneous image characteristics and clinical settings. OOD detection relies on estimating a classification model's uncertainty, which should increase for OOD samples. While existing research has largely focused on uncertainty quantification methods, this work investigates the impact of the classification task itself. Through experiments with eight uncertainty quantification methods across four classification tasks, we demonstrate that OOD detection performance significantly varies with the task, and that the best task depends on the defined ID-OOD criteria; specifically, whether the OOD sample is due to: i) an image characteristic shift or ii) an anatomical feature shift. Furthermore, we reveal that superior OOD detection does not guarantee optimal abstained prediction, underscoring the necessity to align task selection and uncertainty strategies with the specific downstream application in medical image analysis.

Miguel López-Pérez, Marco Miani, Valery Naranjo et al.

Generative models are popular for medical imaging tasks such as anomaly detection, feature extraction, data visualization, or image generation. Since they are parameterized by deep learning models, they are often sensitive to distribution shifts and unreliable when applied to out-of-distribution data, creating a risk of, e.g. underrepresentation bias. This behavior can be flagged using uncertainty quantification methods for generative models, but their availability remains limited. We propose SLUG: A new UQ method for VAEs that combines recent advances in Laplace approximations with stochastic trace estimators to scale gracefully with image dimensionality. We show that our UQ score -- unlike the VAE's encoder variances -- correlates strongly with reconstruction error and racial underrepresentation bias for dermatological images. We also show how pixel-wise uncertainty can detect out-of-distribution image content such as ink, rulers, and patches, which is known to induce learning shortcuts in predictive models.

Vincent Olesen, Nina Weng, Aasa Feragen et al.

Machine learning models have achieved high overall accuracy in medical image analysis. However, performance disparities on specific patient groups pose challenges to their clinical utility, safety, and fairness. This can affect known patient groups - such as those based on sex, age, or disease subtype - as well as previously unknown and unlabeled groups. Furthermore, the root cause of such observed performance disparities is often challenging to uncover, hindering mitigation efforts. In this paper, to address these issues, we leverage Slice Discovery Methods (SDMs) to identify interpretable underperforming subsets of data and formulate hypotheses regarding the cause of observed performance disparities. We introduce a novel SDM and apply it in a case study on the classification of pneumothorax and atelectasis from chest x-rays. Our study demonstrates the effectiveness of SDMs in hypothesis formulation and yields an explanation of previously observed but unexplained performance disparities between male and female patients in widely used chest X-ray datasets and models. Our findings indicate shortcut learning in both classification tasks, through the presence of chest drains and ECG wires, respectively. Sex-based differences in the prevalence of these shortcut features appear to cause the observed classification performance gap, representing a previously underappreciated interaction between shortcut learning and model fairness analyses.

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.

Manxi Lin, Aasa Feragen

Standard spatial convolutions assume input data with a regular neighborhood structure. Existing methods typically generalize convolution to the irregular point cloud domain by fixing a regular "view" through e.g. a fixed neighborhood size, where the convolution kernel size remains the same for each point. However, since point clouds are not as structured as images, the fixed neighbor number gives an unfortunate inductive bias. We present a novel graph convolution named Difference Graph Convolution (diffConv), which does not rely on a regular view. diffConv operates on spatially-varying and density-dilated neighborhoods, which are further adapted by a learned masked attention mechanism. Experiments show that our model is very robust to the noise, obtaining state-of-the-art performance in 3D shape classification and scene understanding tasks, along with a faster inference speed.

Steffen Czolbe, Aasa Feragen, Oswin Krause

Geometric alignment appears in a variety of applications, ranging from domain adaptation, optimal transport, and normalizing flows in machine learning; optical flow and learned augmentation in computer vision and deformable registration within biomedical imaging. A recurring challenge is the alignment of domains whose topology is not the same; a problem that is routinely ignored, potentially introducing bias in downstream analysis. As a first step towards solving such alignment problems, we propose an unsupervised algorithm for the detection of changes in image topology. The model is based on a conditional variational auto-encoder and detects topological changes between two images during the registration step. We account for both topological changes in the image under spatial variation and unexpected transformations. Our approach is validated on two tasks and datasets: detection of topological changes in microscopy images of cells, and unsupervised anomaly detection brain imaging.

Guan Wang, Francois Bernard Lauze, Aasa Feragen

We present a graph neural network model for solving graph-to-graph learning problems. Most deep learning on graphs considers ``simple'' problems such as graph classification or regressing real-valued graph properties. For such tasks, the main requirement for intermediate representations of the data is to maintain the structure needed for output, i.e., keeping classes separated or maintaining the order indicated by the regressor. However, a number of learning tasks, such as regressing graph-valued output, generative models, or graph autoencoders, aim to predict a graph-structured output. In order to successfully do this, the learned representations need to preserve far more structure. We present a conditional auto-regressive model for graph-to-graph learning and illustrate its representational capabilities via experiments on challenging subgraph predictions from graph algorithmics; as a graph autoencoder for reconstruction and visualization; and on pretraining representations that allow graph classification with limited labeled data.

Kasra Arnavaz, Oswin Krause, Kilian Zepf et al.

Motivated by the challenging segmentation task of pancreatic tubular networks, this paper tackles two commonly encountered problems in biomedical imaging: Topological consistency of the segmentation, and expensive or difficult annotation. Our contributions are the following: a) We propose a topological score which measures both topological and geometric consistency between the predicted and ground truth segmentations, applied to model selection and validation. b) We provide a full deep-learning methodology for this difficult noisy task on time-series image data. In our method, we first use a semisupervised U-net architecture, applicable to generic segmentation tasks, which jointly trains an autoencoder and a segmentation network. We then use tracking of loops over time to further improve the predicted topology. This semi-supervised approach allows us to utilize unannotated data to learn feature representations that generalize to test data with high variability, in spite of our annotated training data having very limited variation. Our contributions are validated on a challenging segmentation task, locating tubular structures in the fetal pancreas from noisy live imaging confocal microscopy. We show that our semi-supervised model outperforms not only fully supervised and pre-trained models but also an approach which takes topological consistency into account during training. Further, our approach achieves a mean loop score of 0.808 for detecting loops in the fetal pancreas, compared to a U-net trained with clDice with mean loop score 0.762.

Steffen Czolbe, Oswin Krause, Aasa Feragen

We propose a semantic similarity metric for image registration. Existing metrics like Euclidean Distance or Normalized Cross-Correlation focus on aligning intensity values, giving difficulties with low intensity contrast or noise. Our approach learns dataset-specific features that drive the optimization of a learning-based registration model. We train both an unsupervised approach using an auto-encoder, and a semi-supervised approach using supplemental segmentation data to extract semantic features for image registration. Comparing to existing methods across multiple image modalities and applications, we achieve consistently high registration accuracy. A learned invariance to noise gives smoother transformations on low-quality images.

Steffen Czolbe, Kasra Arnavaz, Oswin Krause et al.

Probabilistic image segmentation encodes varying prediction confidence and inherent ambiguity in the segmentation problem. While different probabilistic segmentation models are designed to capture different aspects of segmentation uncertainty and ambiguity, these modelling differences are rarely discussed in the context of applications of uncertainty. We consider two common use cases of segmentation uncertainty, namely assessment of segmentation quality and active learning. We consider four established strategies for probabilistic segmentation, discuss their modelling capabilities, and investigate their performance in these two tasks. We find that for all models and both tasks, returned uncertainty correlates positively with segmentation error, but does not prove to be useful for active learning.

Steffen Czolbe, Oswin Krause, Aasa Feragen

We propose a semantic similarity metric for image registration. Existing metrics like euclidean distance or normalized cross-correlation focus on aligning intensity values, giving difficulties with low intensity contrast or noise. Our semantic approach learns dataset-specific features that drive the optimization of a learning-based registration model. Comparing to existing unsupervised and supervised methods across multiple image modalities and applications, we achieve consistently high registration accuracy and faster convergence than state of the art, and the learned invariance to noise gives smoother transformations on low-quality images.

M. Jorge Cardoso, Aasa Feragen, Ben Glocker et al.

This compendium gathers all the accepted extended abstracts from the Second International Conference on Medical Imaging with Deep Learning (MIDL 2019), held in London, UK, 8-10 July 2019. Note that only accepted extended abstracts are listed here, the Proceedings of the MIDL 2019 Full Paper Track are published as Volume 102 of the Proceedings of Machine Learning Research (PMLR) http://proceedings.mlr.press/v102/.

Anton Mallasto, Tom Dela Haije, Aasa Feragen

In optimization, the natural gradient method is well-known for likelihood maximization. The method uses the Kullback-Leibler divergence, corresponding infinitesimally to the Fisher-Rao metric, which is pulled back to the parameter space of a family of probability distributions. This way, gradients with respect to the parameters respect the Fisher-Rao geometry of the space of distributions, which might differ vastly from the standard Euclidean geometry of the parameter space, often leading to faster convergence. However, when minimizing an arbitrary similarity measure between distributions, it is generally unclear which metric to use. We provide a general framework that, given a similarity measure, derives a metric for the natural gradient. We then discuss connections between the natural gradient method and multiple other optimization techniques in the literature. Finally, we provide computations of the formal natural gradient to show overlap with well-known cases and to compute natural gradients in novel frameworks.

Anton Mallasto, Jes Frellsen, Wouter Boomsma et al.

Generative Adversial Networks (GANs) have made a major impact in computer vision and machine learning as generative models. Wasserstein GANs (WGANs) brought Optimal Transport (OT) theory into GANs, by minimizing the $1$-Wasserstein distance between model and data distributions as their objective function. Since then, WGANs have gained considerable interest due to their stability and theoretical framework. We contribute to the WGAN literature by introducing the family of $(q,p)$-Wasserstein GANs, which allow the use of more general $p$-Wasserstein metrics for $p\geq 1$ in the GAN learning procedure. While the method is able to incorporate any cost function as the ground metric, we focus on studying the $l^q$ metrics for $q\geq 1$. This is a notable generalization as in the WGAN literature the OT distances are commonly based on the $l^2$ ground metric. We demonstrate the effect of different $p$-Wasserstein distances in two toy examples. Furthermore, we show that the ground metric does make a difference, by comparing different $(q,p)$ pairs on the MNIST and CIFAR-10 datasets. Our experiments demonstrate that changing the ground metric and $p$ can notably improve on the common $(q,p) = (2,1)$ case.

Rune Kok Nielsen, Andreas Nugaard Holm, Aasa Feragen

We study the effect of structural variation in graph data on the predictive performance of graph kernels. To this end, we introduce a novel, noise-robust adaptation of the GraphHopper kernel and validate it on benchmark data, obtaining modestly improved predictive performance on a range of datasets. Next, we investigate the performance of the state-of-the-art Weisfeiler-Lehman graph kernel under increasing synthetic structural errors and find that the effect of introducing errors depends strongly on the dataset.

Anton Mallasto, Søren Hauberg, Aasa Feragen

Latent variable models (LVMs) learn probabilistic models of data manifolds lying in an \emph{ambient} Euclidean space. In a number of applications, a priori known spatial constraints can shrink the ambient space into a considerably smaller manifold. Additionally, in these applications the Euclidean geometry might induce a suboptimal similarity measure, which could be improved by choosing a different metric. Euclidean models ignore such information and assign probability mass to data points that can never appear as data, and vastly different likelihoods to points that are similar under the desired metric. We propose the wrapped Gaussian process latent variable model (WGPLVM), that extends Gaussian process latent variable models to take values strictly on a given ambient Riemannian manifold, making the model blind to impossible data points. This allows non-linear, probabilistic inference of low-dimensional Riemannian submanifolds from data. Our evaluation on diverse datasets show that we improve performance on several tasks, including encoding, visualization and uncertainty quantification.