Research Group Daniel Rückert

In medical imaging, generative models are increasingly relied upon for two distinct but equally critical tasks: reconstruction, where the goal is to restore medical imaging (usually inverse problems like inpainting or superresolution), and generation, where synthetic data is created to augment datasets or carry out counterfactual analysis. Despite shared architecture and learning frameworks, they prioritize different goals: generation seeks high perceptual quality and diversity, while reconstruction focuses on data fidelity and faithfulness. In this work, we introduce a ‘generative model zoo’ and systematically analyze how modern latent diffusion models and autoregressive models navigate the reconstruction-generation spectrum. We benchmark a suite of generative models across representative cardiac medical imaging tasks, focusing on image inpainting with varying masking ratios and sampling strategies, as well as unconditional image generation. Our findings show that diffusion models offer superior perceptual quality for unconditional generation but tend to hallucinate as masking ratios increase, whereas autoregressive models maintain stable perceptual performance across masking levels, albeit with generally lower fidelity.

Vessels are complex structures in the body that have been studied extensively in multiple representations. While voxelization is the most common of them, meshes and parametric models are critical in various applications due to their desirable properties. However, these representations are typically extracted through segmentations and used disjointly from each other. We propose a framework that joins the three representations under differentiable transformations. By leveraging differentiable voxelization, we automatically extract a parametric shape model of the vessels through shape-to-segmentation fitting, where we learn shape parameters from segmentations without the explicit need for ground-truth shape parameters. The vessel is parametrized as centerlines and radii using cubic B-splines, ensuring smoothness and continuity by construction. Meshes are differentiably extracted from the learned shape parameters, resulting in high-fidelity meshes that can be manipulated post-fit. Our method can accurately capture the geometry of complex vessels, as demonstrated by the volumetric fits in experiments on aortas, aneurysms, and brain vessels.

Butterfly tumors are a distinct class of gliomas that span the corpus callosum, producing a characteristic butterfly-shaped appearance on MRI. The distinctive growth pattern of these tumors highlights how white matter fibers and structural connectivity influence brain tumor cell migration. To investigate this relation, we applied biophysical tumor growth models to a large patient cohort, systematically comparing models that incorporate fiber tract information with those that do not. Our results demonstrate that including fiber orientation data significantly improves model accuracy, particularly for a subset of butterfly tumors. These findings highlight the critical role of white matter architecture in tumor spread and suggest that integrating fiber tract information can enhance the precision of radiotherapy target volume delineation.

Supervised contrastive learning (SupCon) has proven to be a powerful alternative to the standard cross-entropy loss for classification of multi-class balanced datasets. However, it struggles to learn well-conditioned representations of datasets with long-tailed class distributions. This problem is potentially exacerbated for binary imbalanced distributions, which are commonly encountered during many real-world problems such as medical diagnosis. In experiments on seven binary datasets of natural and medical images, we show that the performance of SupCon decreases with increasing class imbalance. To substantiate these findings, we introduce two novel metrics that evaluate the quality of the learned representation space. By measuring the class distribution in local neighborhoods, we are able to uncover structural deficiencies of the representation space that classical metrics cannot detect. Informed by these insights, we propose two new supervised contrastive learning strategies tailored to binary imbalanced datasets that improve the structure of the representation space and increase downstream classification accuracy over standard SupCon by up to 35%. We make our code available.

We address the challenge of ensuring user-level DP when individuals contribute varying numbers of data records to a dataset. While group privacy can be used to aggregate record-level budgets, it can be overly pessimistic and lacks flexibility when users contribute varying numbers of data points. We propose a method for accounting for arbitrary numbers of records per user while maintaining a fixed per-user privacy guarantee by leveraging individual privacy assignment. Experimentally, our method yields excellent utility comparable to record-level DP while providing a more meaningful/interpretable protection.

Schizophrenia is a psychiatric disorder hypothesized to result from disturbed brain connectivity. Structural covariance networks (SCN) describe the shared variation in morphological properties emerging from coordinated neurodevelopmental processes and may, thus, be a promising diagnostic biomarker for schizophrenia.We compared the diagnostic value of two SCN computation methods derived from regional gray matter volume (GMV) in 154 patients with a diagnosis of first episode psychosis or recurrent schizophrenia (PAT) and 366 healthy control individuals (HC). The first method (REF-SCN) quantifies the contribution of an individual to a normative reference group’s SCN, and the second approach (KLS-SCN) uses a symmetric version of Kulback-Leibler divergence. Their diagnostic value compared to regional GMV was assessed in a stepwise analysis using a series of linear support vector machines within a nested cross-validation framework and stacked generalization, all models were externally validated in an independent sample (NPAT=71, NHC=74), SCN feature importance was assessed, and the derived risk scores were analyzed for differential relationships with clinical variables.We found that models trained on SCNs were able to classify patients with schizophrenia and combining SCNs and regional GMV in a stacked model improved training (balanced accuracy (BAC)=69.96%) and external validation performance (BAC=67.10%). Among all unimodal models, the highest discovery sample performance was achieved by a model trained on REF-SCN (balanced accuracy (BAC=67.03%). All model decisions were driven by widespread structural covariance alterations involving the somato-motor, default mode, control, visual, and the ventral attention networks. Risk estimates derived from KLS-SCNs and regional GMV, but not REF-SCNs, could be predicted from clinical variables, especially driven by body mass index (BMI) and affect-related negative symptoms. These patterns of results show that different SCN computation approaches capture different aspects of the disease. While REF-SCNs contain valuable information for discriminating schizophrenia from healthy control individuals, KLS-SCNs may capture more nuanced symptom-level characteristics similar to those captured by PCA of regional GMV.

Body mass index (BMI) does not account for substantial inter-individual differences in regional fat and muscle compartments, which are relevant for the prevalence of cardiometabolic and cancer conditions. We applied a validated deep learning pipeline for automated segmentation of whole-body MRI scans in 45,851 adults from the UK Biobank and German National Cohort, enabling harmonized quantification of visceral (VAT), gluteofemoral (GFAT), and abdominal subcutaneous adipose tissue (ASAT), liver fat fraction (LFF), and trunk muscle volume. Associations with clinical conditions were evaluated using compartment measures adjusted for age, sex, height, and BMI. Our analysis demonstrates that regional adiposity and muscle volume show distinct associations with cardiometabolic and cancer prevalence, and that substantial disease heterogeneity exists within BMI strata. The analytic framework and reference data presented here will support future risk stratification efforts and facilitate the integration of automated MRI phenotyping into large-scale population and clinical research.

An electrocardiogram (ECG) is a widely used, cost-effective tool for detecting electrical abnormalities in the heart. However, it cannot directly measure functional parameters, such as ventricular volumes and ejection fraction, which are crucial for assessing cardiac function. Cardiac magnetic resonance (CMR) is the gold standard for these measurements, providing detailed structural and functional insights, but is expensive and less accessible. To bridge this gap, we propose PTACL (Patient and Temporal Alignment Contrastive Learning), a multimodal contrastive learning framework that enhances ECG representations by integrating spatio-temporal information from CMR. PTACL uses global patient-level contrastive loss and local temporal-level contrastive loss. The global loss aligns patient-level representations by pulling ECG and CMR embeddings from the same patient closer together, while pushing apart embeddings from different patients. Local loss enforces fine-grained temporal alignment within each patient by contrasting encoded ECG segments with corresponding encoded CMR frames. This approach enriches ECG representations with diagnostic information beyond electrical activity and transfers more insights between modalities than global alignment alone, all without introducing new learnable weights. We evaluate PTACL on paired ECG-CMR data from 27,951 subjects in the UK Biobank. Compared to baseline approaches, PTACL achieves better performance in two clinically relevant tasks: (1) retrieving patients with similar cardiac phenotypes and (2) predicting CMR-derived cardiac function parameters, such as ventricular volumes and ejection fraction. Our results highlight the potential of PTACL to enhance non-invasive cardiac diagnostics using ECG.

Anatomical atlases are widely used for population studies and analysis. Conditional atlases target a specific sub-population defined via certain conditions, such as demographics or pathologies, and allow for the investigation of fine-grained anatomical differences like morphological changes associated with ageing or disease. Existing approaches use either registration-based methods that are often unable to handle large anatomical variations or generative adversarial models, which are challenging to train since they can suffer from training instabilities. Instead of generating atlases directly in as intensities, we propose using latent diffusion models to generate deformation fields, which transform a general population atlas into one representing a specific sub-population. Our approach ensures structural integrity, enhances interpretability and avoids hallucinations that may arise during direct image synthesis by generating this deformation field and regularising it using a neighbourhood of images. We compare our method to several state-of-the-art atlas generation methods using brain MR images from the UK Biobank. Our method generates highly realistic atlases with smooth transformations and high anatomical fidelity, outperforming existing baselines. We demonstrate the quality of these atlases through comprehensive evaluations, including quantitative metrics for anatomical accuracy, perceptual similarity, and qualitative analyses displaying the consistency and realism of the generated atlases.

Topological correctness, i.e., the preservation of structural integrity and specific characteristics of shape, is a fundamental requirement for medical imaging tasks, such as neuron or vessel segmentation. Despite the recent surge in topology-aware methods addressing this challenge, their real-world applicability is hindered by flawed benchmarking practices. In this paper, we identify critical pitfalls in model evaluation that include inadequate connectivity choices, overlooked topological artifacts in ground truth annotations, and inappropriate use of evaluation metrics. Through detailed empirical analysis, we uncover these issues’ profound impact on the evaluation and ranking of segmentation methods. Drawing from our findings, we propose a set of actionable recommendations to establish fair and robust evaluation standards for topology-aware medical image segmentation methods.

Model multiplicity refers to the existence of multiple machine learning models that describe the data equally well but may produce different predictions on individual samples. In medicine, these models can admit conflicting predictions for the same patient – a risk that is poorly understood and insufficiently addressed.
In this study, we empirically analyze the extent, drivers, and ramifications of predictive multiplicity across diverse medical tasks and model architectures, and show that even small ensembles can mitigate/eliminate predictive multiplicity in practice. Our analysis reveals that (1) standard validation metrics fail to identify a uniquely optimal model and (2) a substantial amount of predictions hinges on arbitrary choices made during model development. Using multiple models instead of a single model reveals instances where predictions differ across equally plausible models – highlighting patients that would receive arbitrary diagnoses if any single model were used. In contrast, (3) a small ensemble paired with an abstention strategy can effectively mitigate measurable predictive multiplicity in practice; predictions with high inter-model consensus may thus be amenable to automated classification. While accuracy is not a principled antidote to predictive multiplicity, we find that (4) higher accuracy achieved through increased model capacity reduces predictive multiplicity. Our findings underscore the clinical importance of accounting for model multiplicity and advocate for ensemble-based strategies to improve diagnostic reliability. In cases where models fail to reach sufficient consensus, we recommend deferring decisions to expert review.

Current AI frameworks for brain decoding and encoding, typically train and test models within the same datasets. This limits their utility for brain computer interfaces (BCI) or neurofeedback, for which it would be useful to pool experiences across individuals to better simulate stimuli not sampled during training. A key obstacle to model generalisation is the degree of variability of inter-subject cortical organisation, which makes it difficult to align or compare cortical signals across participants. In this paper we address this through the use of surface vision transformers, which build a generalisable model of cortical functional dynamics, through encoding the topography of cortical networks and their interactions as a moving image across a surface. This is then combined with tri-modal self-supervised contrastive (CLIP) alignment of audio, video, and fMRI modalities to enable the retrieval of visual and auditory stimuli from patterns of cortical activity (and vice-versa). We validate our approach on 7T task-fMRI data from 174 healthy participants engaged in the movie-watching experiment from the Human Connectome Project (HCP). Results show that it is possible to detect which movie clips an individual is watching purely from their brain activity, even for individuals and movies not seen during training. Further analysis of attention maps reveals that our model captures individual patterns of brain activity that reflect semantic and visual systems. This opens the door to future personalised simulations of brain function.

Accurately estimating the informativeness of individual samples in a dataset is an important objective in deep learning, as it can guide sample selection, which can improve model efficiency and accuracy by removing redundant or potentially harmful samples. We propose Laplace Sample Information (LSI) measure of sample informativeness grounded in information theory widely applicable across model architectures and learning settings. LSI leverages a Bayesian approximation to the weight posterior and the KL divergence to measure the change in the parameter distribution induced by a sample of interest from the dataset. We experimentally show that LSI is effective in ordering the data with respect to typicality, detecting mislabeled samples, measuring class-wise informativeness, and assessing dataset difficulty. We demonstrate these capabilities of LSI on image and text data in supervised and unsupervised settings. Moreover, we show that LSI can be computed efficiently through probes and transfers well to the training of large models.

Topological correctness plays a critical role in many image segmentation tasks, yet most networks are trained using pixel-wise loss functions, such as Dice, neglecting topological accuracy. Existing topology-aware methods often lack robust topological guarantees, are limited to specific use cases, or impose high computational costs. In this work, we propose a novel, graph-based framework for topologically accurate image segmentation that is both computationally efficient and generally applicable. Our method constructs a component graph that fully encodes the topological information of both the prediction and ground truth, allowing us to efficiently identify topologically critical regions and aggregate a loss based on local neighborhood information. Furthermore, we introduce a strict topological metric capturing the homotopy equivalence between the union and intersection of prediction-label pairs. We formally prove the topological guarantees of our approach and empirically validate its effectiveness on binary and multi-class datasets. Our loss demonstrates state-of-the-art performance with up to fivefold faster loss computation compared to persistent homology methods.

Cardiovascular diseases (CVD) can be diagnosed using various diagnostic modalities. The electrocardiogram (ECG) is a cost-effective and widely available diagnostic aid that provides functional information of the heart. However, its ability to classify and spatially localise CVD is limited. In contrast, cardiac magnetic resonance (CMR) imaging provides detailed structural information of the heart and thus enables evidence-based diagnosis of CVD, but long scan times and high costs limit its use in clinical routine. In this work, we present a deep learning strategy for cost-effective and comprehensive cardiac screening solely from ECG. Our approach combines multimodal contrastive learning with masked data modelling to transfer domain-specific information from CMR imaging to ECG representations. In extensive experiments using data from 40,044 UK Biobank subjects, we demonstrate the utility and generalisability of our method for subject-specific risk prediction of CVD and the prediction of cardiac phenotypes using only ECG data. Specifically, our novel multimodal pre-training paradigm improves performance by up to 12.19% for risk prediction and 27.59% for phenotype prediction. In a qualitative analysis, we demonstrate that our learned ECG representations incorporate information from CMR image regions of interest.

Data reconstruction attacks on machine learning models pose a substantial threat to privacy, potentially leaking sensitive information. Although defending against such attacks using differential privacy (DP) provides theoretical guarantees, determining appropriate DP parameters remains challenging. Current formal guarantees on the success of data reconstruction suffer from overly stringent assumptions regarding adversary knowledge about the target data, particularly in the image domain, raising questions about their real-world applicability. In this work, we empirically investigate this discrepancy by introducing a reconstruction attack based on diffusion models (DMs) that only assumes adversary access to real-world image priors and specifically targets the DP defense. We find that (1) real-world data priors significantly influence reconstruction success, (2) current reconstruction bounds do not model the risk posed by data priors well, and (3) DMs can serve as heuristic auditing tools for visualizing privacy leakage.

In this paper, we propose a methodology for extracting molecular tumor biomarkers from hyperspectral imaging (HSI), an emerging technology for intraoperative tissue assessment. To achieve this, we employ spectral unmixing, allowing to decompose the spectral signals recorded by the HSI camera into their constituent molecular components. Traditional unmixing approaches are based on physical models that establish a relationship between tissue molecules and the recorded spectra. However, these methods commonly assume a linear relationship between the spectra and molecular content, which does not capture the whole complexity of light-matter interaction. To address this limitation, we introduce a novel unmixing procedure that allows to take into account non-linear optical effects while preserving the computational benefits of linear spectral unmixing. We validate our methodology on an in-vivo brain tissue HSI dataset and demonstrate that the extracted molecular information leads to superior classification performance.

Evaluating deformable image registration (DIR) is challenging due to the inherent trade-off between achieving high alignment accuracy and maintaining deformation regularity. In this work, we introduce a novel evaluation scheme based on the alignment-regularity characteristic (ARC) to systematically capture and analyze this trade-off. We first introduce the ARC curves, which describe the performance of a given registration algorithm as a spectrum measured by alignment and regularity metrics. We further adopt a HyperNetwork-based approach that learns to continuously interpolate across the full regularization range, accelerating the construction and improving the sample density of ARC curves. We empirically demonstrate our evaluation scheme using representative learning-based deformable image registration methods with various network architectures and transformation models on two public datasets. We present a range of findings not evident from existing evaluation practices and provide general recommendations for model evaluation and selection using our evaluation scheme. All code relevant is made publicly available.

The histopathological classification of whole-slide images (WSIs) is a fundamental task in digital pathology; yet it requires extensive time and expertise from specialists. While deep learning methods show promising results, they typically process WSIs by dividing them into artificial patches, which inherently prevents a network from learning from the entire image context, disregards natural tissue structures and compromises interpretability. Our method overcomes this limitation through a novel graph-based framework that constructs WSI graph representations. The WSI-graph efficiently captures essential histopathological information in a compact form. We build tissue representations (nodes) that follow biological boundaries rather than arbitrary patches all while providing interpretable features for explainability. Through adaptive graph coarsening guided by learned embeddings, we progressively merge regions while maintaining discriminative local features and enabling efficient global information exchange. In our method’s final step, we solve the diagnostic task through a graph attention network. We empirically demonstrate strong performance on multiple challenging tasks such as cancer stage classification and survival prediction, while also identifying predictive factors using Integrated Gradients.

Direct image-to-graph transformation is a challenging task that involves solving object detection and relationship prediction in a single model. Due to this task’s complexity, large training datasets are rare in many domains, making the training of deep-learning methods challenging. This data sparsity necessitates transfer learning strategies akin to the state-of-the-art in general computer vision. In this work, we introduce a set of methods enabling cross-domain and cross-dimension learning for image-to-graph transformers. We propose (1) a regularized edge sampling loss to effectively learn object relations in multiple domains with different numbers of edges, (2) a domain adaptation framework for image-to-graph transformers aligning image- and graph-level features from different domains, and (3) a projection function that allows using 2D data for training 3D transformers. We demonstrate our method’s utility in cross-domain and cross-dimension experiments, where we utilize labeled data from 2D road networks for simultaneous learning in vastly different target domains. Our method consistently outperforms standard transfer learning and self-supervised pretraining on challenging benchmarks, such as retinal or whole-brain vessel graph extraction.

Normative representation learning focuses on understanding the typical anatomical distributions from large datasets of medical scans from healthy individuals. Generative Artificial Intelligence (AI) leverages this attribute to synthesize images that accurately reflect these normative patterns. This capability enables the AI allowing them to effectively detect and correct anomalies in new, unseen pathological data without the need for expert labeling. Traditional anomaly detection methods often evaluate the anomaly detection performance, overlooking the crucial role of normative learning. In our analysis, we introduce novel metrics, specifically designed to evaluate this facet in AI models. We apply these metrics across various generative AI frameworks, including advanced diffusion models, and rigorously test them against complex and diverse brain pathologies. In addition, we conduct a large multi-reader study to compare these metrics to experts’ evaluations. Our analysis demonstrates that models proficient in normative learning exhibit exceptional versatility, adeptly detecting a wide range of unseen medical conditions.

The success of foundation models in natural language processing and computer vision has motivated similar approaches for general time series analysis. While these models are effective for a variety of tasks, their applicability in medical domains with limited data remains largely unexplored. To address this, we investigate the effectiveness of foundation models in medical time series analysis involving electroencephalography (EEG). Through extensive experiments on tasks such as age prediction, seizure detection, and the classification of clinically relevant EEG events, we compare their diagnostic accuracy with that of specialised EEG models. Our analysis shows that foundation models extract meaningful EEG features, outperform specialised models even without domain adaptation, and localise task-specific biomarkers. Moreover, we demonstrate that diagnostic accuracy is substantially influenced by architectural choices such as context length. Overall, our study reveals that foundation models with general time series understanding eliminate the dependency on large domain-specific datasets, making them valuable tools for clinical practice.

Dark-field radiography of the human chest has been demonstrated to have promising potential for the analysis of the lung microstructure and the diagnosis of respiratory diseases. However, previous studies of dark-field chest radiographs evaluated the lung signal only in the inspiratory breathing state. Our work aims to add a new perspective to these previous assessments by locally comparing dark-field lung information between different respiratory states. To this end, we discuss suitable image registration methods for dark-field chest radiographs to enable consistent spatial alignment of the lung in distinct breathing states. Utilizing full inspiration and expiration scans from a clinical chronic obstructive pulmonary disease study, we assess the performance of the proposed registration framework and outline applicable evaluation approaches. Our regional characterization of lung dark-field signal changes between the breathing states provides a proof-of-principle that dynamic radiography-based lung function assessment approaches may benefit from considering registered dark-field images in addition to standard plain chest radiographs.

Glioblastoma, a highly aggressive brain tumor, poses major challenges due to its poor prognosis and high morbidity rates. Partial differential equation-based models offer promising potential to enhance therapeutic outcomes by simulating patient-specific tumor behavior for improved radiotherapy planning. However, model calibration remains a bottleneck due to the high computational demands of optimization methods like Monte Carlo sampling and evolutionary algorithms. To address this, we recently introduced an approach leveraging a neural forward solver with gradient-based optimization to significantly reduce calibration time. This approach requires a highly accurate and fully differentiable forward model. We investigate multiple architectures, including (i) an enhanced TumorSurrogate, (ii) a modified nnU-Net, and (iii) a 3D Vision Transformer (ViT). The optimized TumorSurrogate achieved the best overall results, excelling in both tumor outline matching and voxel-level prediction of tumor cell concentration. It halved the MSE relative to the baseline model and achieved the highest Dice score across all tumor cell concentration thresholds. Our study demonstrates significant enhancement in forward solver performance and outlines important future research directions.

Predicting future brain states is crucial for understanding healthy aging and neurodegenerative diseases. Longitudinal brain MRI registration, a cornerstone for such analyses, has long been limited by its inability to forecast future developments, reliance on extensive, dense longitudinal data, and the need to balance registration accuracy with temporal smoothness. In this work, we present emph{TimeFlow}, a novel framework for longitudinal brain MRI registration that overcomes all these challenges. Leveraging a U-Net architecture with temporal conditioning inspired by diffusion models, TimeFlow enables accurate longitudinal registration and facilitates prospective analyses through future image prediction. Unlike traditional methods that depend on explicit smoothness regularizers and dense sequential data, TimeFlow achieves temporal consistency and continuity without these constraints. Experimental results highlight its superior performance in both future timepoint prediction and registration accuracy compared to state-of-the-art methods. Additionally, TimeFlow supports novel biological brain aging analyses, effectively differentiating neurodegenerative conditions from healthy aging. It eliminates the need for segmentation, thereby avoiding the challenges of non-trivial annotation and inconsistent segmentation errors. TimeFlow paves the way for accurate, data-efficient, and annotation-free prospective analyses of brain aging and chronic diseases.

Image registration is fundamental in medical imaging applications, such as disease progression analysis or radiation therapy planning. The primary objective of image registration is to precisely capture the deformation between two or more images, typically achieved by minimizing an optimization problem. Due to its inherent ill-posedness, regularization is a key component in driving the solution toward anatomically meaningful deformations. A wide range of regularization methods has been proposed for both conventional and deep learning-based registration. However, the appropriate application of regularization techniques often depends on the specific registration problem, and no one-fits-all method exists. Despite its importance, regularization is often overlooked or addressed with default approaches, assuming existing methods are sufficient. A comprehensive and structured review remains missing. This review addresses this gap by introducing a novel taxonomy that systematically categorizes the diverse range of proposed regularization methods. It highlights the emerging field of learned regularization, which leverages data-driven techniques to automatically derive deformation properties from the data. Moreover, this review examines the transfer of regularization methods from conventional to learning-based registration, identifies open challenges, and outlines future research directions. By emphasizing the critical role of regularization in image registration, we hope to inspire the research community to reconsider regularization strategies in modern registration algorithms and to explore this rapidly evolving field further.

Biophysical modeling of brain tumors has emerged as a promising strategy for personalizing radiotherapy planning by estimating the otherwise hidden distribution of tumor cells within the brain. However, many existing state-of-the-art methods are computationally intensive, limiting their widespread translation into clinical practice. In this work, we propose an efficient and direct method that utilizes soft physical constraints to estimate the tumor cell concentration from preoperative MRI of brain tumor patients. Our approach optimizes a 3D tumor concentration field by simultaneously minimizing the difference between the observed MRI and a physically informed loss function. Compared to existing state-of-the-art techniques, our method significantly improves predicting tumor recurrence on two public datasets with a total of 192 patients while maintaining a clinically viable runtime of under one minute - a substantial reduction from the 30 minutes required by the current best approach. Furthermore, we showcase the generalizability of our framework by incorporating additional imaging information and physical constraints, highlighting its potential to translate to various medical diffusion phenomena with imperfect data.

Feature attribution methods attempt to explain neural network predictions by identifying relevant features. However, establishing a cohesive framework for assessing feature attribution remains a challenge. There are several views through which we can evaluate attributions. One principal lens is to observe the effect of perturbing attributed features on the model’s behavior (i.e., faithfulness). While providing useful insights, existing faithfulness evaluations suffer from shortcomings that we reveal in this paper. In this work, we propose two new perspectives within the faithfulness paradigm that reveal intuitive properties: soundness and completeness. Soundness assesses the degree to which attributed features are truly predictive features, while completeness examines how well the resulting attribution reveals all the predictive features. The two perspectives are based on a firm mathematical foundation and provide quantitative metrics that are computable through efficient algorithms. We apply these metrics to mainstream attribution methods, offering a novel lens through which to analyze and compare feature attribution methods.

Employing pre-trained Large Language Models (LLMs) has become the de facto standard in Natural Language Processing (NLP) despite their extensive data requirements. Motivated by the recent surge in research focused on training LLMs with limited data, particularly in low-resource domains and languages, this paper surveys recent transfer learning approaches to optimize model performance in downstream tasks where data is scarce. We first address initial and continued pre-training strategies to better leverage prior knowledge in unseen domains and languages. We then examine how to maximize the utility of limited data during fine-tuning and few-shot learning. The final section takes a task-specific perspective, reviewing models and methods suited for different levels of data scarcity. Our goal is to provide practitioners with practical guidelines for overcoming the challenges posed by constrained data while also highlighting promising directions for future research.

Federated learning enhanced with Differential Privacy (DP) is a powerful privacy-preserving strategy to protect individuals sharing their sensitive data for processing in fields such as medicine and healthcare. Many medical applications, for example magnetic resonance imaging (MRI), rely on complex-valued signal processing techniques for data acquisition and analysis. However, the appropriate application of DP to complex-valued data is still underexplored. To address this issue, from the theoretical side, we introduce the complex-valued Gaussian mechanism, whose behaviour we characterise in terms of f-DP, -DP and Rényi-DP. Moreover, we generalise the fundamental algorithm DP stochastic gradient descent to complex-valued neural networks and present novel complex-valued neural network primitives compatible with DP. Experimentally, we showcase a proof-of-concept by training federated complex-valued neural networks with DP on a real-world task (MRI pulse sequence classification in k-space), yielding excellent utility and privacy. Our results highlight the relevance of combining federated learning with robust privacy-preserving techniques in the MRI context.

Image-to-graph transformers can effectively encode image information in graphs but are typically difficult to train and require large annotated datasets. Contrastive learning can increase data efficiency by enhancing feature representations, but existing methods are not applicable to graph labels because they operate on categorical label spaces. In this work, we propose a method enabling supervised contrastive learning for image-to-graph transformers. We introduce two supervised contrastive loss formulations based on graph similarity between label pairs that we approximate using a graph neural network. Our approach avoids tailored data augmentation techniques and can be easily integrated into existing training pipelines. We perform multiple empirical studies showcasing performance improvements across various metrics.

Interpretability, particularly in terms of human understandable concepts, is essential for building trust in machine learning models for disease classification. However, state-of-the-art image classifiers exhibit limited interpretability, posing a significant barrier to their acceptance in clinical practice. To address this, our work introduces two graph representations of the retinal vasculature, aiming to bridge the gap between high-performance classifiers and human-understandable interpretability concepts in ophthalmology. We use these graphs with the aim of training graph neural networks (GNNs) for disease staging. First, we formally and experimentally show that GNNs can learn known clinical biomarkers. In that, we show that GNNs can learn human interpretable concepts. Next, we train GNNs for disease staging and study how different aggregation strategies lead the GNN to learn more and less human interpretable features. Finally, we propose a visualization for integrated gradients on graphs, which allows us to identify if GNN models have learned human-understandable representations of the data.

Topological accuracy in medical image segmentation is a highly important property for downstream applications such as network analysis and flow modeling in vessels or cell counting. Recently, significant methodological advancements have brought well-founded concepts from algebraic topology to binary segmentation. However, these approaches have been underexplored in multi-class segmentation scenarios, where topological errors are common. We propose a general loss function for topologically faithful multi-class segmentation extending the recent Betti matching concept, which is based on induced matchings of persistence barcodes. We project the N-class segmentation problem to N single-class segmentation tasks, which allows us to use 1-parameter persistent homology, making training of neural networks computationally feasible. We validate our method on a comprehensive set of four medical datasets with highly variant topological characteristics. Our loss formulation significantly enhances topological correctness in cardiac, cell, artery-vein, and Circle of Willis segmentation.

This research explores the integration of language models and unsupervised anomaly detection in medical imaging, addressing two key questions: (1) Can language models enhance the interpretability of anomaly detection maps? and (2) Can anomaly maps improve the generalizability of language models in open-set anomaly detection tasks? To investigate these questions, we introduce a new dataset for multi-image visual question-answering on brain magnetic resonance images encompassing multiple conditions. We propose KQ-Former (Knowledge Querying Transformer), which is designed to optimally align visual and textual information in limited-sample contexts. Our model achieves a 60.81% accuracy on closed questions, covering disease classification and severity across 15 different classes. For open questions, KQ-Former demonstrates a 70% improvement over the baseline with a BLEU-4 score of 0.41, and achieves the highest entailment ratios (up to 71.9%) and lowest contradiction ratios (down to 10.0%) among various natural language inference models. Furthermore, integrating anomaly maps results in an 18% accuracy increase in detecting open-set anomalies, thereby enhancing the language model’s generalizability to previously unseen medical conditions.

VoxelMorph, proposed in 2018, utilizes Convolutional Neural Networks (CNNs) to address medical image registration problems. In 2021 TransMorph advanced this approach by replacing CNNs with Attention mechanisms, claiming enhanced performance. More recently, the rise of Mamba with selective state space models has led to MambaMorph, which substituted Attention with Mamba blocks, asserting superior registration. These developments prompt a critical question: does chasing the latest computational trends with “more advanced” computational blocks genuinely enhance registration accuracy, or is it merely hype? Furthermore, the role of classic high-level registration-specific designs, such as coarse-to-fine pyramid mechanism, correlation calculation, and iterative optimization, warrants scrutiny, particularly in differentiating their influence from the aforementioned low-level computational blocks. In this study, we critically examine these questions through a rigorous evaluation in brain MRI registration. We employed modularized components for each block and ensured unbiased comparisons across all methods and designs to disentangle their effects on performance. Our findings indicate that adopting “advanced” computational elements fails to significantly improve registration accuracy. Instead, well-established registration-specific designs offer fair improvements, enhancing results by a marginal 1.5% over the baseline. Our findings emphasize the importance of rigorous, unbiased evaluation and contribution disentanglement of all low- and high-level registration components, rather than simply following the computer vision trends with “more advanced” computational blocks. We advocate for simpler yet effective solutions and novel evaluation metrics that go beyond conventional registration accuracy, warranting further research across various organs and modalities.

General vision encoders like DINOv2 and SAM have recently transformed computer vision. Even though they are trained on natural images, such encoder models have excelled in medical imaging, e.g., in classification, segmentation, and registration. However, no in-depth comparison of different state-of-the-art general vision encoders for medical registration is available. In this work, we investigate how well general vision encoder features can be used in the dissimilarity metrics for medical image registration. We explore two encoders that were trained on natural images as well as one that was fine-tuned on medical data. We apply the features within the well-established B-spline FFD registration framework. In extensive experiments on cardiac cine MRI data, we find that using features as additional guidance for conventional metrics improves the registration quality.

Report generation models offer fine-grained textual interpretations of medical images like chest X-rays, yet they often lack interactivity (i.e. the ability to steer the generation process through user queries) and localized interpretability (i.e. visually grounding their predictions), which we deem essential for future adoption in clinical practice. While there have been efforts to tackle these issues, they are either limited in their interactivity by not supporting textual queries or fail to also offer localized interpretability. Therefore, we propose a novel multitask architecture and training paradigm integrating textual prompts and bounding boxes for diverse aspects like anatomical regions and pathologies. We call this approach the Chest X-Ray Explainer (ChEX). Evaluations across a heterogeneous set of 9 chest X-ray tasks, including localized image interpretation and report generation, showcase its competitiveness with SOTA models while additional analysis demonstrates ChEX’s interactive capabilities.

The rapid development of artificial intelligence (AI) has gained importance, with many tools already entering our daily lives. The medical field of radiation oncology is also subject to this development, with AI entering all steps of the patient journey. In this review article, we summarize contemporary AI techniques and explore the clinical applications of AI-based automated segmentation models in radiotherapy planning, focusing on delineation of organs at risk (OARs), the gross tumor volume (GTV), and the clinical target volume (CTV). Emphasizing the need for precise and individualized plans, we review various commercial and freeware segmentation tools and also state-of-the-art approaches. Through our own findings and based on the literature, we demonstrate improved efficiency and consistency as well as time savings in different clinical scenarios. Despite challenges in clinical implementation such as domain shifts, the potential benefits for personalized treatment planning are substantial. The integration of mathematical tumor growth models and AI-based tumor detection further enhances the possibilities for refining target volumes. As advancements continue, the prospect of one-stop-shop segmentation and radiotherapy planning represents an exciting frontier in radiotherapy, potentially enabling fast treatment with enhanced precision and individualization.

In differentially private (DP) machine learning, the privacy guarantees of DP mechanisms are often reported and compared on the basis of a single pε,δq-pair. This practice overlooks that DP guarantees can vary substantially even between mechanisms sharing a given pε,δq, and potentially introduces privacy vulnerabilities which can remain undetected. This motivates the need for robust, rigorous methods for comparing DP guarantees in such cases. Here, we introduce the ∆-divergence between mechanisms which quantifies the worst-case excess privacy vulnerability of choosing one mechanism over another in terms of pε,δq, f-DP and in terms of a newly presented Bayesian interpretation. Moreover, as a generalisation of the Blackwell theorem, it is endowed with strong decision-theoretic foundations. Through application examples, we show that our techniques can facilitate informed decision-making and reveal gaps in the current understanding of privacy risks, as current practices in DP-SGD often result in choosing mechanisms with high excess privacy vulnerabilities.

The automation of chest X-ray reporting has garnered significant interest due to the time-consuming nature of the task. However, the clinical accuracy of free-text reports has proven challenging to quantify using natural language processing metrics, given the complexity of medical information, the variety of writing styles, and the potential for typos and inconsistencies. Structured reporting and standardized reports, on the other hand, can provide consistency and formalize the evaluation of clinical correctness. However, high-quality annotations for structured reporting are scarce. Therefore, we propose a method to predict clinical findings defined by sentences in structured reporting templates, which can be used to fill such templates. The approach involves training a contrastive language-image model using chest X-rays and related free-text radiological reports, then creating textual prompts for each structured finding and optimizing a classifier to predict clinical findings in the medical image. Results show that even with limited image-level annotations for training, the method can accomplish the structured reporting tasks of severity assessment of cardiomegaly and localizing pathologies in chest X-rays.

Background: Alterations of the superficial retinal vasculature are commonly observed in multiple sclerosis (MS) and can be visualized through optical coherence tomography angiography (OCTA).
Objectives: This study aimed to examine changes in the retinal vasculature during MS and to integrate findings into current concepts of the underlying pathology.
Methods: In this cross-sectional study, including 259 relapsing–remitting MS patients and 78 healthy controls, we analyzed OCTAs using deep-learning-based segmentation algorithm tools.
Results: We identified a loss of small-sized vessels (diameter < 10 µm) in the superficial vascular complex in all MS eyes, irrespective of their optic neuritis (ON) history. This alteration was associated with MS disease burden and appears independent of retinal ganglion cell loss. In contrast, an observed reduction of medium-sized vessels (diameter 10–20 µm) was specific to eyes with a history of ON and was closely linked to ganglion cell atrophy.
Conclusion: These findings suggest distinct atrophy patterns in retinal vessels in patients with MS. Further studies are necessary to investigate retinal vessel alterations and their underlying pathology in MS.

Artificial intelligence (AI) models are vulnerable to information leakage of their training data, which can be highly sensitive, for example, in medical imaging. Privacy-enhancing technologies, such as differential privacy (DP), aim to circumvent these susceptibilities. DP is the strongest possible protection for training models while bounding the risks of inferring the inclusion of training samples or reconstructing the original data. DP achieves this by setting a quantifiable privacy budget. Although a lower budget decreases the risk of information leakage, it typically also reduces the performance of such models. This imposes a trade-off between robust performance and stringent privacy. Additionally, the interpretation of a privacy budget remains abstract and challenging to contextualize. Here we contrast the performance of artificial intelligence models at various privacy budgets against both theoretical risk bounds and empirical success of reconstruction attacks. We show that using very large privacy budgets can render reconstruction attacks impossible, while drops in performance are negligible. We thus conclude that not using DP at all is negligent when applying artificial intelligence models to sensitive data. We deem our results to lay a foundation for further debates on striking a balance between privacy risks and model performance.

Cardiac magnetic resonance (CMR) image acquisition requires subjects to hold their breath while 2D cine images are acquired. This process assumes that the heart remains in the same position across all slices. However, differences in breathhold positions or patient motion introduce 3D slice misalignments. In this work, we propose an algorithm that simultaneously aligns all SA and LA slices by maximizing the pair-wise intensity agreement between their intersections. Unlike previous works, our approach is formulated as a subject-specific optimization problem and requires no prior knowledge of the underlying anatomy. We quantitatively demonstrate that the proposed method is robust against a large range of rotations and translations by synthetically misaligning 10 motion-free datasets and aligning them back using the proposed method.

The prevailing deep learning-based methods of predicting cardiac segmentation involve reconstructed magnetic resonance (MR) images. The heavy dependency of segmentation approaches on image quality significantly limits the acceleration rate in fast MR reconstruction. Moreover, the practice of treating reconstruction and segmentation as separate sequential processes leads to artifact generation and information loss in the intermediate stage. These issues pose a great risk to achieving high-quality outcomes. To leverage the redundant k-space information overlooked in this dual-step pipeline, we introduce a novel approach to directly deriving segmentations from sparse k-space samples using a transformer (DiSK). DiSK operates by globally extracting latent features from 2D+time k-space data with attention blocks and subsequently predicting the segmentation label of query points. We evaluate our model under various acceleration factors (ranging from 4 to 64) and compare against two image-based segmentation baselines. Our model consistently outperforms the baselines in Dice and Hausdorff distances across foreground classes for all presented sampling rates.

Motivation: High acceleration factors place a limit on MRI image reconstruction. This limit is extended to segmentation models when treating these as subsequent independent processes.
Goal(s): Our goal is to produce segmentations directly from sparse k-space measurements without the need for intermediate image reconstruction.
Approach: We employ a transformer architecture to encode global k-space information into latent features. The produced latent vectors condition queried coordinates during decoding to generate segmentation class probabilities.
Results: The model is able to produce better segmentations across high acceleration factors than image-based segmentation baselines.
Impact: Cardiac segmentation directly from undersampled k-space samples circumvents the need for an intermediate image reconstruction step. This allows the potential to assess myocardial structure and function on higher acceleration factors than methods that rely on images as input.

Background: Artificial intelligence (AI) models are increasingly used in the medical domain. However, as medical data is highly sensitive, special precautions to ensure its protection are required. The gold standard for privacy preservation is the introduction of differential privacy (DP) to model training. Prior work indicates that DP has negative implications on model accuracy and fairness, which are unacceptable in medicine and represent a main barrier to the widespread use of privacy-preserving techniques. In this work, we evaluated the effect of privacy-preserving training of AI models regarding accuracy and fairness compared to non-private training.
Methods: We used two datasets: (1) A large dataset (N=193,311) of high quality clinical chest radiographs, and (2) a dataset (N=1625) of 3D abdominal computed tomography (CT) images, with the task of classifying the presence of pancreatic ductal adenocarcinoma (PDAC). Both were retrospectively collected and manually labeled by experienced radiologists. We then compared non-private deep convolutional neural networks (CNNs) and privacy-preserving (DP) models with respect to privacy-utility trade-offs measured as area under the receiver operating characteristic curve (AUROC), and privacy-fairness trade-offs, measured as Pearson’s r or Statistical Parity Difference.
Results: We find that, while the privacy-preserving training yields lower accuracy, it largely does not amplify discrimination against age, sex or co-morbidity. However, we find an indication that difficult diagnoses and subgroups suffer stronger performance hits in private training.
Conclusions: Our study shows that – under the challenging realistic circumstances of a real-life clinical dataset – the privacy-preserving training of diagnostic deep learning models is possible with excellent diagnostic accuracy and fairness.

Optical coherence tomography angiography (OCTA) is a non-invasive imaging modality that can acquire high-resolution volumes of the retinal vasculature and aid the diagnosis of ocular, neurological and cardiac diseases. Segmenting the visible blood vessels is a common first step when extracting quantitative biomarkers from these images. Classical segmentation algorithms based on thresholding are strongly affected by image artifacts and limited signal-to-noise ratio. The use of modern, deep learning-based segmentation methods has been inhibited by a lack of large datasets with detailed annotations of the blood vessels. To address this issue, recent work has employed transfer learning, where a segmentation network is trained on synthetic OCTA images and is then applied to real data. However, the previously proposed simulations fail to faithfully model the retinal vasculature and do not provide effective domain adaptation. Because of this, current methods are unable to fully segment the retinal vasculature, in particular the smallest capillaries. In this work, we present a lightweight simulation of the retinal vascular network based on space colonization for faster and more realistic OCTA synthesis. We then introduce three contrast adaptation pipelines to decrease the domain gap between real and artificial images. We demonstrate the superior segmentation performance of our approach in extensive quantitative and qualitative experiments on three public datasets that compare our method to traditional computer vision algorithms and supervised training using human annotations. Finally, we make our entire pipeline publicly available, including the source code, pretrained models, and a large dataset of synthetic OCTA images.

Differentially private mechanisms restrict the membership inference capabilities of powerful (optimal) adversaries against machine learning models. Such adversaries are rarely encountered in practice. In this work, we examine a more realistic threat model relaxation, where (sub-optimal) adversaries lack access to the exact model training database, but may possess related or partial data. We then formally characterise and experimentally validate adversarial membership inference capabilities in this setting in terms of hypothesis testing errors. Our work helps users to interpret the privacy properties of sensitive data processing systems under realistic threat model relaxations and choose appropriate noise levels for their use-case.

Feature attribution explains neural network outputs by identifying relevant input features. How do we know if the identified features are indeed relevant to the network? This notion is referred to as faithfulness, an essential property that reflects the alignment between the identified (attributed) features and the features used by the model. One recent trend to test faithfulness is to design the data such that we know which input features are relevant to the label and then train a model on the designed data. Subsequently, the identified features are evaluated by comparing them with these designed ground truth features. However, this idea has the underlying assumption that the neural network learns to use all and only these designed features, while there is no guarantee that the learning process trains the network in this way. In this paper, we solve this missing link by explicitly designing the neural network by manually setting its weights, along with designing data, so we know precisely which input features in the dataset are relevant to the designed network. Thus, we can test faithfulness in AttributionLab, our designed synthetic environment, which serves as a sanity check and is effective in filtering out attribution methods. If an attribution method is not faithful in a simple controlled environment, it can be unreliable in more complex scenarios. Furthermore, the AttributionLab environment serves as a laboratory for controlled experiments through which we can study feature attribution methods, identify issues, and suggest potential improvements.

The European Commission’s draft for the European Health Data Space (EHDS) aims to empower citizens to access their personal health data and share it with physicians and other health-care providers. It further defines procedures for the secondary use of electronic health data for research and development. Although this planned legislation is undoubtedly a step in the right direction, implementation approaches could potentially result in centralised data silos that pose data privacy and security risks for individuals. To address this concern, we propose federated personal health data spaces, a novel architecture for storing, managing, and sharing personal electronic health records that puts citizens at the centre—both conceptually and technologically. The proposed architecture puts citizens in control by storing personal health data on a combination of personal devices rather than in centralised data silos. We describe how this federated architecture fits within the EHDS and can enable the same features as centralised systems while protecting the privacy of citizens. We further argue that increased privacy and control do not contradict the use of electronic health data for research and development. Instead, data sovereignty and transparency encourage active participation in studies and data sharing. This combination of privacy-by-design and transparent, privacy-preserving data sharing can enable health-care leaders to break the privacy-exploitation barrier, which currently limits the secondary use of health data in many cases.

Image synthesis is increasingly being adopted in medical image processing, for example for data augmentation or inter-modality image translation. In these critical applications, the generated images must fulfill a high standard of biological correctness. A particular requirement for these images is global consistency, i.e an image being overall coherent and structured so that all parts of the image fit together in a realistic and meaningful way. Yet, established image quality metrics do not explicitly quantify this property of synthetic images. In this work, we introduce two metrics that can measure the global consistency of synthetic images on a per-image basis. To measure the global consistency, we presume that a realistic image exhibits consistent properties, e.g., a person’s body fat in a whole-body MRI, throughout the depicted object or scene. Hence, we quantify global consistency by predicting and comparing explicit attributes of images on patches using supervised trained neural networks. Next, we adapt this strategy to an unlabeled setting by measuring the similarity of implicit image features predicted by a self-supervised trained network. Our results demonstrate that predicting explicit attributes of synthetic images on patches can distinguish globally consistent from inconsistent images. Implicit representations of images are less sensitive to assess global consistency but are still serviceable when labeled data is unavailable. Compared to established metrics, such as the FID, our method can explicitly measure global consistency on a per-image basis, enabling a dedicated analysis of the biological plausibility of single synthetic images.

Neural implicit representations (NIRs) enable to generate and parametrize the transformation for image registration in a continuous way. By design, these representations are image-pair-specific, meaning that for each signal a new multi-layer perceptron has to be trained. In this work, we investigate for the first time the potential of existent NIR generalisation methods for image registration and propose novel methods for the registration of a group of image pairs using NIRs. To exploit the generalisation potential of NIRs, we encode the fixed and moving image volumes to latent representations, which are then used to condition or modulate the NIR. Using ablation studies on a 3D benchmark dataset, we show that our methods are able to generalise to a set of image pairs with a performance comparable to pairwise registration using NIRs when trained on and datasets. Our results demonstrate the potential of generalised NIRs for 3D deformable image registration.

Age-related macular degeneration (AMD) is the leading cause of blindness in the elderly. Current grading systems based on imaging biomarkers only coarsely group disease stages into broad categories that lack prognostic value for future disease progression. It is widely believed that this is due to their focus on a single point in time, disregarding the dynamic nature of the disease. In this work, we present the first method to automatically propose biomarkers that capture temporal dynamics of disease progression. Our method represents patient time series as trajectories in a latent feature space built with contrastive learning. Then, individual trajectories are partitioned into atomic sub-sequences that encode transitions between disease states. These are clustered using a newly introduced distance metric. In quantitative experiments we found our method yields temporal biomarkers that are predictive of conversion to late AMD. Furthermore, these clusters were highly interpretable to ophthalmologists who confirmed that many of the clusters represent dynamics that have previously been linked to the progression of AMD, even though they are currently not included in any clinical grading system.

Segmentation of anatomical shapes from medical images has taken an important role in the automation of clinical measurements. While typical deep-learning segmentation approaches are performed on discrete voxels, the underlying objects being analysed exist in a real-valued continuous space. Approaches that rely on convolutional neural networks (CNNs) are limited to grid-like inputs and not easily applicable to sparse or partial measurements. We propose a novel family of image segmentation models that tackle many of CNNs’ shortcomings: Neural Implicit Segmentation Functions (NISF). Our framework takes inspiration from the field of neural implicit functions where a network learns a mapping from a real-valued coordinate-space to a shape representation. NISFs have the ability to segment anatomical shapes in high-dimensional continuous spaces. Training is not limited to voxelized grids, and covers applications with sparse and partial data. Interpolation between observations is learnt naturally in the training procedure and requires no post-processing. Furthermore, NISFs allow the leveraging of learnt shape priors to make predictions for regions outside of the original image plane. We go on to show the framework achieves dice scores of on a (3D+t) short-axis cardiac segmentation task using the UK Biobank dataset. We also provide a qualitative analysis on our frameworks ability to perform segmentation and image interpolation on unseen regions of an image volume at arbitrary resolutions.

The skeleton of a digital image is a compact representation of its topology, geometry, and scale. It has utility in many computer vision applications, such as image description, segmentation, and registration. However, skeletonization has only seen limited use in contemporary deep learning solutions. Most existing skeletonization algorithms are not differentiable, making it impossible to integrate them with gradient-based optimization. Compatible algorithms based on morphological operations and neural networks have been proposed, but their results often deviate from the geometry and topology of the true medial axis. This work introduces the first three-dimensional skeletonization algorithm that is both compatible with gradient-based optimization and preserves an object’s topology. Our method is exclusively based on matrix additions and multiplications, convolutional operations, basic non-linear functions, and sampling from a uniform probability distribution, allowing it to be easily implemented in any major deep learning library. In benchmarking experiments, we prove the advantages of our skeletonization algorithm compared to non-differentiable, morphological, and neural-network-based baselines. Finally, we demonstrate the utility of our algorithm by integrating it with two medical image processing applications that use gradient-based optimization: deep-learning-based blood vessel segmentation, and multimodal registration of the mandible in computed tomography and magnetic resonance images.

Feature attribution is arguably the predominant approach for illuminating black-box neural networks. This dissertation rethinks feature attribution by leveraging critical neural pathways, identifying input features with predictive information, and evaluating feature attribution using the neural network model. The dissertation also rethinks feature attribution for the explanation of medical imaging models.

Do black-box neural network models learn clinically relevant features for fracture diagnosis? The answer not only establishes reliability, quenches scientific curiosity, but also leads to explainable and verbose findings that can assist the radiologists in the final and increase trust. This work identifies the concepts networks use for vertebral fracture diagnosis in CT images. This is achieved by associating concepts to neurons highly correlated with a specific diagnosis in the dataset. The concepts are either associated with neurons by radiologists pre-hoc or are visualized during a specific prediction and left for the user’s interpretation. We evaluate which concepts lead to correct diagnosis and which concepts lead to false positives. The proposed frameworks and analysis pave the way for reliable and explainable vertebral fracture diagnosis.

One challenging property lurking in medical datasets is the imbalanced data distribution, where the frequency of the samples between the different classes is not balanced. Training a model on an imbalanced dataset can introduce unique challenges to the learning problem where a model is biased towards the highly frequent class. Many methods are proposed to tackle the distributional differences and the imbalanced problem. However, the impact of these approaches on the learned features is not well studied. In this paper, we look deeper into the internal units of neural networks to observe how handling data imbalance affects the learned features. We study several popular cost-sensitive approaches for handling data imbalance and analyze the feature maps of the convolutional neural networks from multiple perspectives: analyzing the alignment of salient features with pathologies and analyzing the pathology-related concepts encoded by the networks. Our study reveals differences and insights regarding the trained models that are not reflected by quantitative metrics such as AUROC and AP and show up only by looking at the models through a lens.

It is a mystery which input features contribute to a neural network’s output. Various explanation (feature attribution) methods are proposed in the literature to shed light on the problem. One peculiar observation is that these explanations (attributions) point to different features as being important. The phenomenon raises the question, which explanation to trust? We propose a framework for evaluating the explanations using the neural network model itself. The framework leverages the network to generate input features that impose a particular behavior on the output. Using the generated features, we devise controlled experimental setups to evaluate whether an explanation method conforms to an axiom. Thus we propose an empirical framework for axiomatic evaluation of explanation methods. We evaluate well-known and promising explanation solutions using the proposed framework. The framework provides a toolset to reveal properties and drawbacks within existing and future explanation solutions

The automation of chest X-ray reporting has garnered significant interest due to the time-consuming nature of the task. However, the clinical accuracy of free-text reports has proven challenging to quantify using natural language processing metrics, given the complexity of medical information, the variety of writing styles, and the potential for typos and inconsistencies. Structured reporting and standardized reports, on the other hand, can provide consistency and formalize the evaluation of clinical correctness. However, high-quality annotations for structured reporting are scarce. Therefore, we propose a method to predict clinical findings defined by sentences in structured reporting templates, which can be used to fill such templates. The approach involves training a contrastive language-image model using chest X-rays and related free-text radiological reports, then creating textual prompts for each structured finding and optimizing a classifier to predict clinical findings in the medical image. Results show that even with limited image-level annotations for training, the method can accomplish the structured reporting tasks of severity assessment of cardiomegaly and localizing pathologies in chest X-rays.

One principal approach for illuminating a black-box neural network is feature attribution, i.e. identifying the importance of input features for the network’s prediction. The predictive information of features is recently proposed as a proxy for the measure of their importance. So far, the predictive information is only identified for latent features by placing an information bottleneck within the network. We propose a method to identify features with predictive information in the input domain. The method results in fine-grained identification of input features’ information and is agnostic to network architecture. The core idea of our method is leveraging a bottleneck on the input that only lets input features associated with predictive latent features pass through. We compare our method with several feature attribution methods using mainstream feature attribution evaluation experiments. The code is publicly available.

Convolutional neural networks are showing promise in the automatic diagnosis of thoracic pathologies on chest x-rays. Their black-box nature has sparked many recent works to explain the prediction via input feature attribution methods (aka saliency methods). However, input feature attribution methods merely identify the importance of input regions for the prediction and lack semantic interpretation of model behavior. In this work, we first identify the semantics associated with internal units (feature maps) of the network. We proceed to investigate the following questions; Does a regression model that is only trained with COVID-19 severity scores implicitly learn visual patterns associated with thoracic pathologies? Does a network that is trained on weakly labeled data (e.g. healthy, unhealthy) implicitly learn pathologies? Moreover, we investigate the effect of pretraining and data imbalance on the interpretability of learned features. In addition to the analysis, we propose semantic attribution to semantically explain each prediction. We present our findings using publicly available chest pathologies (CheXpert [5], NIH ChestX-ray8 [25]) and COVID-19 datasets (BrixIA [20], and COVID-19 chest X-ray segmentation dataset [4]).

Neural networks have demonstrated remarkable performance in classification and regression tasks on chest X-rays. In order to establish trust in the clinical routine, the networks’ prediction mechanism needs to be interpretable. One principal approach to interpretation is feature attribution. Feature attribution methods identify the importance of input features for the output prediction. Building on Information Bottleneck Attribution (IBA) method, for each prediction we identify the chest X-ray regions that have high mutual information with the network’s output. Original IBA identifies input regions that have sufficient predictive information. We propose Inverse IBA to identify all informative regions. Thus all predictive cues for pathologies are highlighted on the X-rays, a desirable property for chest X-ray diagnosis. Moreover, we propose Regression IBA for explaining regression models. Using Regression IBA we observe that a model trained on cumulative severity score labels implicitly learns the severity of different X-ray regions. Finally, we propose Multi-layer IBA to generate higher resolution and more detailed attribution/saliency maps. We evaluate our methods using both human-centric (ground-truth-based) interpretability metrics, and human-agnostic feature importance metrics on NIH Chest X-ray8 and BrixIA datasets.

Is critical input information encoded in specific sparse pathways within the neural network? In this work, we discuss the problem of identifying these critical pathways and subsequently leverage them for interpreting the network’s response to an input. The pruning objective — selecting the smallest group of neurons for which the response remains equivalent to the original network — has been previously proposed for identifying critical pathways. We demonstrate that sparse pathways derived from pruning do not necessarily encode critical input information. To ensure sparse pathways include critical fragments of the encoded input information, we propose pathway selection via neurons’ contribution to the response. We proceed to explain how critical pathways can reveal critical input features. We prove that pathways selected via neuron contribution are locally linear (in an ℓ 2 -ball), a property that we use for proposing a feature attribution method: ‘pathway gradient’. We validate our interpretation method using mainstream evaluation experiments. The validation of pathway gradient interpretation method further confirms that selected pathways using neuron contributions correspond to critical input features. The code 1 2 is publicly available.

Segmentation of Multiple Sclerosis (MS) lesions in longitudinal brain MR scans is performed for monitoring the progression of MS lesions. We hypothesize that the spatio-temporal cues in longitudinal data can aid the segmentation algorithm. Therefore, we propose a multi-task learning approach by defining an auxiliary self-supervised task of deformable registration between two time-points to guide the neural network toward learning from spatio-temporal changes. We show the efficacy of our method on a clinical dataset comprised of 70 patients with one follow-up study for each patient. Our results show that spatio-temporal information in longitudinal data is a beneficial cue for improving segmentation. We improve the result of current state-of-the-art by 2.6% in terms of overall score (p < 0.05).