Mads Nielsen

Stefan Sommer, Mads Nielsen, Sune Darkner et al.

To achieve sparse parametrizations that allows intuitive analysis, we aim to represent deformation with a basis containing interpretable elements, and we wish to use elements that have the description capacity to represent the deformation compactly. To accomplish this, we introduce in this paper higher-order momentum distributions in the LDDMM registration framework. While the zeroth order moments previously used in LDDMM only describe local displacement, the first-order momenta that are proposed here represent a basis that allows local description of affine transformations and subsequent compact description of non-translational movement in a globally non-rigid deformation. The resulting representation contains directly interpretable information from both mathematical and modeling perspectives. We develop the mathematical construction of the registration framework with higher-order momenta, we show the implications for sparse image registration and deformation description, and we provide examples of how the parametrization enables registration with a very low number of parameters. The capacity and interpretability of the parametrization using higher-order momenta lead to natural modeling of articulated movement, and the method promises to be useful for quantifying ventricle expansion and progressing atrophy during Alzheimer's disease.

Andreas D. Lauritzen, My Catarina von Euler-Chelpin, Elsebeth Lynge et al.

Purpose: Risk-stratified breast cancer screening might improve early detection and efficiency without comprising quality. However, modern mammography-based risk models do not ensure adaptation across vendor-domains and rely on cancer precursors, associated with short-term risk, which might limit long-term risk assessment. We report a cross-vendor mammographic texture model for long-term risk. Approach: The texture model was robustly trained using two systematically designed case-control datasets. Textural features, indicative of future breast cancer, were learned by excluding samples with diagnosed/potential malignancies from training. An augmentation-based domain adaption technique, based on flavorization of mammographic views, ensured generalization across vendor-domains. The model was validated in 66,607 consecutively screened Danish women with flavorized Siemens views and 25,706 Dutch women with Hologic-processed views. Performances were evaluated for interval cancers (IC) within two years from screening and long-term cancers (LTC) from two years after screening. The texture model was combined with established risk factors to flag 10% of women with the highest risk. Results: In Danish women, the texture model achieved an area under the receiver operating characteristic (AUC) of 0.71 and 0.65 for ICs and LTCs, respectively. In Dutch women with Hologic-processed views, the AUCs were not different from AUCs in Danish women with flavorized views. The AUC for texture combined with established risk factors increased to 0.68 for LTCs. The 10% of women flagged as high-risk accounted for 25.5% of ICs and 24.8% of LTCs. Conclusions: The texture model robustly estimated long-term breast cancer risk while adapting to an unseen processed vendor-domain and identified a clinically relevant high-risk subgroup.

Sebastian Nørgaard Llambias, Julia Machnio, Asbjørn Munk et al.

Medical image analysis using deep learning frameworks has advanced healthcare by automating complex tasks, but many existing frameworks lack flexibility, modularity, and user-friendliness. To address these challenges, we introduce Yucca, an open-source AI framework available at https://github.com/Sllambias/yucca, designed specifically for medical imaging applications and built on PyTorch and PyTorch Lightning. Yucca features a three-tiered architecture: Functional, Modules, and Pipeline, providing a comprehensive and customizable solution. Evaluated across diverse tasks such as cerebral microbleeds detection, white matter hyperintensity segmentation, and hippocampus segmentation, Yucca achieves state-of-the-art results, demonstrating its robustness and versatility. Yucca offers a powerful, flexible, and user-friendly platform for medical image analysis, inviting community contributions to advance its capabilities and impact.

Asbjørn Munk, Jakob Ambsdorf, Sebastian Llambias et al.

This study investigates the impact of self-supervised pretraining of 3D semantic segmentation models on a large-scale, domain-specific dataset. We introduce BRAINS-45K, a dataset of 44,756 brain MRI volumes from public sources, the largest public dataset available, and revisit a number of design choices for pretraining modern segmentation architectures by simplifying and optimizing state-of-the-art methods, and combining them with a novel augmentation strategy. The resulting AMAES framework is based on masked-image-modeling and intensity-based augmentation reversal and balances memory usage, runtime, and finetuning performance. Using the popular U-Net and the recent MedNeXt architecture as backbones, we evaluate the effect of pretraining on three challenging downstream tasks, covering single-sequence, low-resource settings, and out-of-domain generalization. The results highlight that pretraining on the proposed dataset with AMAES significantly improves segmentation performance in the majority of evaluated cases, and that it is beneficial to pretrain the model with augmentations, despite pretraing on a large-scale dataset. Code and model checkpoints for reproducing results, as well as the BRAINS-45K dataset are available at \url{https://github.com/asbjrnmunk/amaes}.

Asbjørn Munk, Stefano Cerri, Vardan Nersesjan et al.

Clinical deployment of automated brain MRI analysis faces a fundamental challenge: clinical data is heterogeneous and noisy, and high-quality labels are prohibitively costly to obtain. Self-supervised learning (SSL) can address this by leveraging the vast amounts of unlabeled data produced in clinical workflows to train robust \textit{foundation models} that adapt out-of-domain with minimal supervision. However, the development of foundation models for brain MRI has been limited by small pretraining datasets and in-domain benchmarking focused on high-quality, research-grade data. To address this gap, we organized the FOMO25 challenge as a satellite event at MICCAI 2025. FOMO25 provided participants with a large pretraining dataset, FOMO60K, and evaluated models on data sourced directly from clinical workflows in few-shot and out-of-domain settings. Tasks covered infarct classification, meningioma segmentation, and brain age regression, and considered both models trained on FOMO60K (method track) and any data (open track). Nineteen foundation models from sixteen teams were evaluated using a standardized containerized pipeline. Results show that (a) self-supervised pretraining improves generalization on clinical data under domain shift, with the strongest models trained \textit{out-of-domain} surpassing supervised baselines trained \textit{in-domain}. (b) No single pretraining objective benefits all tasks: MAE favors segmentation, hybrid reconstruction-contrastive objectives favor classification, and (c) strong performance was achieved by small pretrained models, and improvements from scaling model size and training duration did not yield reliable benefits.

Mostafa Mehdipour Ghazi, Mads Nielsen

Medical images used in clinical practice are heterogeneous and not the same quality as scans studied in academic research. Preprocessing breaks down in extreme cases when anatomy, artifacts, or imaging parameters are unusual or protocols are different. Methods robust to these variations are most needed. A novel deep learning method is proposed for fast and accurate segmentation of the human brain into 132 regions. The proposed model uses an efficient U-Net-like network and benefits from the intersection points of different views and hierarchical relations for the fusion of the orthogonal 2D planes and brain labels during the end-to-end training. Weakly supervised learning is deployed to take the advantage of partially labeled data for the whole brain segmentation and estimation of the intracranial volume (ICV). Moreover, data augmentation is used to expand the magnetic resonance imaging (MRI) data by generating realistic brain scans with high variability for robust training of the model while preserving data privacy. The proposed method can be applied to brain MRI data including skull or any other artifacts without preprocessing the images or a drop in performance. Several experiments using different atlases are conducted to evaluate the segmentation performance of the trained model compared to the state-of-the-art, and the results show higher segmentation accuracy and robustness of the proposed model compared to the existing methods across different intra- and inter-domain datasets.

Sebastian Nørgaard Llambias, Mads Nielsen, Mostafa Mehdipour Ghazi

Deep learning-based models in medical imaging often struggle to generalize effectively to new scans due to data heterogeneity arising from differences in hardware, acquisition parameters, population, and artifacts. This limitation presents a significant challenge in adopting machine learning models for clinical practice. We propose an unsupervised method for robust domain adaptation in brain MRI segmentation by leveraging MRI-specific augmentation techniques. To evaluate the effectiveness of our method, we conduct extensive experiments across diverse datasets, modalities, and segmentation tasks, comparing against the state-of-the-art methods. The results show that our proposed approach achieves high accuracy, exhibits broad applicability, and showcases remarkable robustness against domain shift in various tasks, surpassing the state-of-the-art performance in the majority of cases.

Neus Rodeja Ferrer, Malini Vendela Sagar, Kiril Vadimovic Klein et al.

Cerebral Microbleeds (CMBs), typically captured as hypointensities from susceptibility-weighted imaging (SWI), are particularly important for the study of dementia, cerebrovascular disease, and normal aging. Recent studies on COVID-19 have shown an increase in CMBs of coronavirus cases. Automatic detection of CMBs is challenging due to the small size and amount of CMBs making the classes highly imbalanced, lack of publicly available annotated data, and similarity with CMB mimics such as calcifications, irons, and veins. Hence, the existing deep learning methods are mostly trained on very limited research data and fail to generalize to unseen data with high variability and cannot be used in clinical setups. To this end, we propose an efficient 3D deep learning framework that is actively trained on multi-domain data. Two public datasets assigned for normal aging, stroke, and Alzheimer's disease analysis as well as an in-house dataset for COVID-19 assessment are used to train and evaluate the models. The obtained results show that the proposed method is robust to low-resolution images and achieves 78% recall and 80% precision on the entire test set with an average false positive of 1.6 per scan.

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.

Asbjørn Munk, Ao Ma, Mads Nielsen

The current state-of-the art techniques for image segmentation are often based on U-Net architectures, a U-shaped encoder-decoder networks with skip connections. Despite the powerful performance, the architecture often does not perform well when used on data which has different characteristics than the data it was trained on. Many techniques for improving performance in the presence of domain shift have been developed, however typically only have loose connections to the theory of domain adaption. In this work, we propose an unsupervised domain adaptation framework for U-Nets with theoretical guarantees based on the Margin Disparity Discrepancy [1] called the MDD-UNet. We evaluate the proposed technique on the task of hippocampus segmentation, and find that the MDD-UNet is able to learn features which are domain-invariant with no knowledge about the labels in the target domain. The MDD-UNet improves performance over the standard U-Net on 11 out of 12 combinations of datasets. This work serves as a proof of concept by demonstrating an improvement on the U-Net in it's standard form without modern enhancements, which opens up a new avenue of studying domain adaptation for models with very large hypothesis spaces from both methodological and practical perspectives. Code is available at https://github.com/asbjrnmunk/mdd-unet.

Julia Machnio, Mads Nielsen, Mostafa Mehdipour Ghazi

Active learning (AL) seeks to reduce annotation costs by selecting the most informative samples for labeling, making it particularly valuable in resource-constrained settings. However, traditional evaluation methods, which focus solely on final accuracy, fail to capture the full dynamics of the learning process. To address this gap, we propose PALM (Performance Analysis of Active Learning Models), a unified and interpretable mathematical model that characterizes AL trajectories through four key parameters: achievable accuracy, coverage efficiency, early-stage performance, and scalability. PALM provides a predictive description of AL behavior from partial observations, enabling the estimation of future performance and facilitating principled comparisons across different strategies. We validate PALM through extensive experiments on CIFAR-10/100 and ImageNet-50/100/200, covering a wide range of AL methods and self-supervised embeddings. Our results demonstrate that PALM generalizes effectively across datasets, budgets, and strategies, accurately predicting full learning curves from limited labeled data. Importantly, PALM reveals crucial insights into learning efficiency, data space coverage, and the scalability of AL methods. By enabling the selection of cost-effective strategies and predicting performance under tight budget constraints, PALM lays the basis for more systematic, reproducible, and data-efficient evaluation of AL in both research and real-world applications. The code is available at: https://github.com/juliamachnio/PALM.

Mikkel Odgaard, Kiril Vadimovic Klein, Sanne Møller Thysen et al.

The widespread adoption of Electronic Health Records (EHR) has significantly increased the amount of available healthcare data. This has allowed models inspired by Natural Language Processing (NLP) and Computer Vision, which scale exceptionally well, to be used in EHR research. Particularly, BERT-based models have surged in popularity following the release of BEHRT and Med-BERT. Subsequent models have largely built on these foundations despite the fundamental design choices of these pioneering models remaining underexplored. Through incremental optimization, we study BERT-based EHR modeling and isolate the sources of improvement for key design choices, giving us insights into the effect of data representation, individual technical components, and training procedure. Evaluating this across a set of generic tasks (death, pain treatment, and general infection), we showed that improving data representation can increase the average downstream performance from 0.785 to 0.797 AUROC ($p<10^{-7}$), primarily when including medication and timestamps. Improving the architecture and training protocol on top of this increased average downstream performance to 0.801 AUROC ($p<10^{-7}$). We then demonstrated the consistency of our optimization through a rigorous evaluation across 25 diverse clinical prediction tasks. We observed significant performance increases in 17 out of 25 tasks and improvements in 24 tasks, highlighting the generalizability of our results. Our findings provide a strong foundation for future work and aim to increase the trustworthiness of BERT-based EHR models.

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.

Jon Middleton, Marko Bauer, Kaining Sheng et al.

The identification and localisation of pathological tissues in medical images continues to command much attention among deep learning practitioners. When trained on abundant datasets, deep neural networks can match or exceed human performance. However, the scarcity of annotated data complicates the training of these models. Data augmentation techniques can compensate for a lack of training samples. However, many commonly used augmentation methods can fail to provide meaningful samples during model fitting. We present local gamma augmentation, a technique for introducing new instances of intensities in pathological tissues. We leverage local gamma augmentation to compensate for a bias in intensities corresponding to ischemic stroke lesions in human brain MRIs. On three datasets, we show how local gamma augmentation can improve the image-level sensitivity of a deep neural network tasked with ischemic lesion segmentation on magnetic resonance images.

Asbjørn Munk, Stefano Cerri, Jakob Ambsdorf et al.

We present FOMO60K, a large-scale, heterogeneous dataset of 60,529 brain Magnetic Resonance Imaging (MRI) scans from 13,900 sessions and 11,187 subjects, aggregated from 16 publicly available sources. The dataset includes both clinical- and research-grade images, multiple MRI sequences, and a wide range of anatomical and pathological variability, including scans with large brain anomalies. Minimal preprocessing was applied to preserve the original image characteristics while reducing barriers to entry for new users. Accompanying code for self-supervised pretraining and finetuning is provided. FOMO60K is intended to support the development and benchmarking of self-supervised learning methods in medical imaging at scale.

Julia Machnio, Mads Nielsen, Mostafa Mehdipour Ghazi

White matter hyperintensities (WMH) are key imaging markers in cognitive aging, Alzheimer's disease (AD), and related dementias. Although automated methods for WMH segmentation have advanced, most provide only global lesion load and overlook their spatial distribution across distinct white matter regions. We propose a deep learning framework for robust WMH segmentation and localization, evaluated across public datasets and an independent Alzheimer's Disease Neuroimaging Initiative (ADNI) cohort. Our results show that the predicted lesion loads are in line with the reference WMH estimates, confirming the robustness to variations in lesion load, acquisition, and demographics. Beyond accurate segmentation, we quantify WMH load within anatomically defined regions and combine these measures with brain structure volumes to assess diagnostic value. Regional WMH volumes consistently outperform global lesion burden for disease classification, and integration with brain atrophy metrics further improves performance, reaching area under the curve (AUC) values up to 0.97. Several spatially distinct regions, particularly within anterior white matter tracts, are reproducibly associated with diagnostic status, indicating localized vulnerability in AD. These results highlight the added value of regional WMH quantification. Incorporating localized lesion metrics alongside atrophy markers may enhance early diagnosis and stratification in neurodegenerative disorders.

Julia Machnio, Sebastian Nørgaard Llambias, Mads Nielsen et al.

White matter hyperintensities (WMH) are radiological markers of small vessel disease and neurodegeneration, whose accurate segmentation and spatial localization are crucial for diagnosis and monitoring. While multimodal MRI offers complementary contrasts for detecting and contextualizing WM lesions, existing approaches often lack flexibility in handling missing modalities and fail to integrate anatomical localization efficiently. We propose a deep learning framework for WM lesion segmentation and localization that operates directly in native space using single- and multi-modal MRI inputs. Our study evaluates four input configurations: FLAIR-only, T1-only, concatenated FLAIR and T1, and a modality-interchangeable setup. It further introduces a multi-task model for jointly predicting lesion and anatomical region masks to estimate region-wise lesion burden. Experiments conducted on the MICCAI WMH Segmentation Challenge dataset demonstrate that multimodal input significantly improves the segmentation performance, outperforming unimodal models. While the modality-interchangeable setting trades accuracy for robustness, it enables inference in cases with missing modalities. Joint lesion-region segmentation using multi-task learning was less effective than separate models, suggesting representational conflict between tasks. Our findings highlight the utility of multimodal fusion for accurate and robust WMH analysis, and the potential of joint modeling for integrated predictions.

William Michael Laprade, Jesper Cairo Westergaard, Svend Christensen et al.

Spectral imaging data acquired via multispectral and hyperspectral cameras can have hundreds of channels, where each channel records the reflectance at a specific wavelength and bandwidth. Time and resource constraints limit our ability to collect large spectral datasets, making it difficult to build and train predictive models from scratch. In the RGB domain, we can often alleviate some of the limitations of smaller datasets by using pretrained foundational models as a starting point. However, most existing foundation models are pretrained on large datasets of 3-channel RGB images, severely limiting their effectiveness when used with spectral imaging data. The few spectral foundation models that do exist usually have one of two limitations: (1) they are built and trained only on remote sensing data limiting their application in proximal spectral imaging, (2) they utilize the more widely available multispectral imaging datasets with less than 15 channels restricting their use with hundred-channel hyperspectral images. To alleviate these issues, we propose a large-scale foundational model and dataset built upon the masked autoencoder architecture that takes advantage of spectral channel encoding, spatial-spectral masking and ImageNet pretraining for an adaptable and robust model for downstream spectral imaging tasks.

Jakob Krogh Petersen, Valdemar Licht, Mads Nielsen et al.

Multi-modal models require aligned, shared embedding spaces. However, common CLIP-based approaches need large amounts of samples and do not natively support 3D or tabular data, both of which are crucial in the medical domain. To address these issues, we revisit CLIP-style alignment by training a domain-specific 3D foundation model as an image encoder and demonstrate that modality alignment is feasible with only 62 MRI scans. Our approach is enabled by a simple embedding accumulation strategy required for training in 3D, which scales the amount of negative pairs across batches in order to stabilize training. We perform a thorough evaluation of various design choices, including the choice of backbone and loss functions, and evaluate the proposed methodology on zero-shot classification and image-retrieval tasks. While zero-shot image-retrieval remains challenging, zero-shot classification results demonstrate that the proposed approach can meaningfully align the representations of 3D MRI with tabular data.

Milla E. Nielsen, Mads Nielsen, Mostafa Mehdipour Ghazi

Alzheimer's disease (AD) is the leading cause of dementia, and its early detection is crucial for effective intervention, yet current diagnostic methods often fall short in sensitivity and specificity. This study aims to detect significant indicators of early AD by extracting and integrating various imaging biomarkers, including radiomics, hippocampal texture descriptors, cortical thickness measurements, and deep learning features. We analyze structural magnetic resonance imaging (MRI) scans from the Alzheimer's Disease Neuroimaging Initiative (ADNI) cohorts, utilizing comprehensive image analysis and machine learning techniques. Our results show that combining multiple biomarkers significantly improves detection accuracy. Radiomics and texture features emerged as the most effective predictors for early AD, achieving AUCs of 0.88 and 0.72 for AD and MCI detection, respectively. Although deep learning features proved to be less effective than traditional approaches, incorporating age with other biomarkers notably enhanced MCI detection performance. Additionally, our findings emphasize the continued importance of classical imaging biomarkers in the face of modern deep-learning approaches, providing a robust framework for early AD diagnosis.

Mauricio Orbes-Arteaga, Thomas Varsavsky, Lauge Sorensen et al.

The insertion of deep learning in medical image analysis had lead to the development of state-of-the art strategies in several applications such a disease classification, as well as abnormality detection and segmentation. However, even the most advanced methods require a huge and diverse amount of data to generalize. Because in realistic clinical scenarios, data acquisition and annotation is expensive, deep learning models trained on small and unrepresentative data tend to outperform when deployed in data that differs from the one used for training (e.g data from different scanners). In this work, we proposed a domain adaptation methodology to alleviate this problem in segmentation models. Our approach takes advantage of the properties of adversarial domain adaptation and consistency training to achieve more robust adaptation. Using two datasets with white matter hyperintensities (WMH) annotations, we demonstrated that the proposed method improves model generalization even in corner cases where individual strategies tend to fail.

Stephan Sloth Lorenzen, Christian Igel, Mads Nielsen

The information bottleneck (IB) principle has been suggested as a way to analyze deep neural networks. The learning dynamics are studied by inspecting the mutual information (MI) between the hidden layers and the input and output. Notably, separate fitting and compression phases during training have been reported. This led to some controversy including claims that the observations are not reproducible and strongly dependent on the type of activation function used as well as on the way the MI is estimated. Our study confirms that different ways of binning when computing the MI lead to qualitatively different results, either supporting or refusing IB conjectures. To resolve the controversy, we study the IB principle in settings where MI is non-trivial and can be computed exactly. We monitor the dynamics of quantized neural networks, that is, we discretize the whole deep learning system so that no approximation is required when computing the MI. This allows us to quantify the information flow without measurement errors. In this setting, we observed a fitting phase for all layers and a compression phase for the output layer in all experiments; the compression in the hidden layers was dependent on the type of activation function. Our study shows that the initial IB results were not artifacts of binning when computing the MI. However, the critical claim that the compression phase may not be observed for some networks also holds true.

Mostafa Mehdipour Ghazi, Lauge Sørensen, Sébastien Ourselin et al.

Learning temporal patterns from multivariate longitudinal data is challenging especially in cases when data is sporadic, as often seen in, e.g., healthcare applications where the data can suffer from irregularity and asynchronicity as the time between consecutive data points can vary across features and samples, hindering the application of existing deep learning models that are constructed for complete, evenly spaced data with fixed sequence lengths. In this paper, a novel deep learning-based model is developed for modeling multiple temporal features in sporadic data using an integrated deep learning architecture based on a recurrent neural network (RNN) unit and a continuous-time autoregressive (CAR) model. The proposed model, called CARRNN, uses a generalized discrete-time autoregressive model that is trainable end-to-end using neural networks modulated by time lags to describe the changes caused by the irregularity and asynchronicity. It is applied to multivariate time-series regression tasks using data provided for Alzheimer's disease progression modeling and intensive care unit (ICU) mortality rate prediction, where the proposed model based on a gated recurrent unit (GRU) achieves the lowest prediction errors among the proposed RNN-based models and state-of-the-art methods using GRUs and long short-term memory (LSTM) networks in their architecture.

Svetlana Kutuzova, Oswin Krause, Douglas McCloskey et al.

Multimodal generative models should be able to learn a meaningful latent representation that enables a coherent joint generation of all modalities (e.g., images and text). Many applications also require the ability to accurately sample modalities conditioned on observations of a subset of the modalities. Often not all modalities may be observed for all training data points, so semi-supervised learning should be possible. In this study, we propose a novel product-of-experts (PoE) based variational autoencoder that have these desired properties. We benchmark it against a mixture-of-experts (MoE) approach and an approach of combining the modalities with an additional encoder network. An empirical evaluation shows that the PoE based models can outperform the contrasted models. Our experiments support the intuition that PoE models are more suited for a conjunctive combination of modalities.

Raghavendra Selvan, Erik B. Dam, Nicki S. Detlefsen et al.

Pulmonary opacification is the inflammation in the lungs caused by many respiratory ailments, including the novel corona virus disease 2019 (COVID-19). Chest X-rays (CXRs) with such opacifications render regions of lungs imperceptible, making it difficult to perform automated image analysis on them. In this work, we focus on segmenting lungs from such abnormal CXRs as part of a pipeline aimed at automated risk scoring of COVID-19 from CXRs. We treat the high opacity regions as missing data and present a modified CNN-based image segmentation network that utilizes a deep generative model for data imputation. We train this model on normal CXRs with extensive data augmentation and demonstrate the usefulness of this model to extend to cases with extreme abnormalities.

Mostafa Mehdipour Ghazi, Mads Nielsen, Akshay Pai et al.

Weight initialization is important for faster convergence and stability of deep neural networks training. In this paper, a robust initialization method is developed to address the training instability in long short-term memory (LSTM) networks. It is based on a normalized random initialization of the network weights that aims at preserving the variance of the network input and output in the same range. The method is applied to standard LSTMs for univariate time series regression and to LSTMs robust to missing values for multivariate disease progression modeling. The results show that in all cases, the proposed initialization method outperforms the state-of-the-art initialization techniques in terms of training convergence and generalization performance of the obtained solution.

Mauricio Orbes-Arteaga, Thomas Varsavsky, Carole H. Sudre et al.

Supervised learning algorithms trained on medical images will often fail to generalize across changes in acquisition parameters. Recent work in domain adaptation addresses this challenge and successfully leverages labeled data in a source domain to perform well on an unlabeled target domain. Inspired by recent work in semi-supervised learning we introduce a novel method to adapt from one source domain to $n$ target domains (as long as there is paired data covering all domains). Our multi-domain adaptation method utilises a consistency loss combined with adversarial learning. We provide results on white matter lesion hyperintensity segmentation from brain MRIs using the MICCAI 2017 challenge data as the source domain and two target domains. The proposed method significantly outperforms other domain adaptation baselines.

Mauricio Orbes-Arteaga, Jorge Cardoso, Lauge Sørensen et al.

In the absence of sufficient data variation (e.g., scanner and protocol variability) in annotated data, deep neural networks (DNNs) tend to overfit during training. As a result, their performance is significantly lower on data from unseen sources compared to the performance on data from the same source as the training data. Semi-supervised domain adaptation methods can alleviate this problem by tuning networks to new target domains without the need for annotated data from these domains. Adversarial domain adaptation (ADA) methods are a popular choice that aim to train networks in such a way that the features generated are domain agnostic. However, these methods require careful dataset-specific selection of hyperparameters such as the complexity of the discriminator in order to achieve a reasonable performance. We propose to use knowledge distillation (KD) -- an efficient way of transferring knowledge between different DNNs -- for semi-supervised domain adaption of DNNs. It does not require dataset-specific hyperparameter tuning, making it generally applicable. The proposed method is compared to ADA for segmentation of white matter hyperintensities (WMH) in magnetic resonance imaging (MRI) scans generated by scanners that are not a part of the training set. Compared with both the baseline DNN (trained on source domain only and without any adaption to target domain) and with using ADA for semi-supervised domain adaptation, the proposed method achieves significantly higher WMH dice scores.

Mostafa Mehdipour Ghazi, Mads Nielsen, Akshay Pai et al.

Quantitative characterization of disease progression using longitudinal data can provide long-term predictions for the pathological stages of individuals. This work studies the robust modeling of Alzheimer's disease progression using parametric methods. The proposed method linearly maps the individual's age to a disease progression score (DPS) and jointly fits constrained generalized logistic functions to the longitudinal dynamics of biomarkers as functions of the DPS using M-estimation. Robustness of the estimates is quantified using bootstrapping via Monte Carlo resampling, and the estimated inflection points of the fitted functions are used to temporally order the modeled biomarkers in the disease course. Kernel density estimation is applied to the obtained DPSs for clinical status classification using a Bayesian classifier. Different M-estimators and logistic functions, including a novel type proposed in this study, called modified Stannard, are evaluated on the data from the Alzheimer's Disease Neuroimaging Initiative (ADNI) for robust modeling of volumetric MRI and PET biomarkers, CSF measurements, as well as cognitive tests. The results show that the modified Stannard function fitted using the logistic loss achieves the best modeling performance with an average normalized MAE of 0.991 across all biomarkers and bootstraps. Applied to the ADNI test set, this model achieves a multiclass AUC of 0.934 in clinical status classification. The obtained results for the proposed model outperform almost all state-of-the-art results in predicting biomarker values and classifying clinical status. Finally, the experiments show that the proposed model, trained using abundant ADNI data, generalizes well to data from the National Alzheimer's Coordinating Center (NACC) with an average normalized MAE of 1.182 and a multiclass AUC of 0.929.

Mostafa Mehdipour Ghazi, Mads Nielsen, Akshay Pai et al.

Disease progression modeling (DPM) using longitudinal data is a challenging machine learning task. Existing DPM algorithms neglect temporal dependencies among measurements, make parametric assumptions about biomarker trajectories, do not model multiple biomarkers jointly, and need an alignment of subjects' trajectories. In this paper, recurrent neural networks (RNNs) are utilized to address these issues. However, in many cases, longitudinal cohorts contain incomplete data, which hinders the application of standard RNNs and requires a pre-processing step such as imputation of the missing values. Instead, we propose a generalized training rule for the most widely used RNN architecture, long short-term memory (LSTM) networks, that can handle both missing predictor and target values. The proposed LSTM algorithm is applied to model the progression of Alzheimer's disease (AD) using six volumetric magnetic resonance imaging (MRI) biomarkers, i.e., volumes of ventricles, hippocampus, whole brain, fusiform, middle temporal gyrus, and entorhinal cortex, and it is compared to standard LSTM networks with data imputation and a parametric, regression-based DPM method. The results show that the proposed algorithm achieves a significantly lower mean absolute error (MAE) than the alternatives with p < 0.05 using Wilcoxon signed rank test in predicting values of almost all of the MRI biomarkers. Moreover, a linear discriminant analysis (LDA) classifier applied to the predicted biomarker values produces a significantly larger AUC of 0.90 vs. at most 0.84 with p < 0.001 using McNemar's test for clinical diagnosis of AD. Inspection of MAE curves as a function of the amount of missing data reveals that the proposed LSTM algorithm achieves the best performance up until more than 74% missing values. Finally, it is illustrated how the method can successfully be applied to data with varying time intervals.

Mauricio Orbes Arteaga, Lauge Sørensen, M. Jorge Cardoso et al.

For proper generalization performance of convolutional neural networks (CNNs) in medical image segmentation, the learnt features should be invariant under particular non-linear shape variations of the input. To induce invariance in CNNs to such transformations, we propose Probabilistic Augmentation of Data using Diffeomorphic Image Transformation (PADDIT) -- a systematic framework for generating realistic transformations that can be used to augment data for training CNNs. We show that CNNs trained with PADDIT outperforms CNNs trained without augmentation and with generic augmentation in segmenting white matter hyperintensities from T1 and FLAIR brain MRI scans.

Francois Lauze, Mads Nielsen

We develop in this paper a generic Bayesian framework for the joint estimation of motion and recovery of missing data in a damaged video sequence. Using standard maximum a posteriori to variational formulation rationale, we derive generic minimum energy formulations for the estimation of a reconstructed sequence as well as motion recovery. We instantiate these energy formulations and from their Euler-Lagrange Equations, we propose a full multiresolution algorithms in order to compute good local minimizers for our energies and discuss their numerical implementations, focusing on the missing data recovery part, i.e. inpainting. Experimental results for synthetic as well as real sequences are presented. Image sequences and extra material is available at http://image.diku.dk/francois/seqinp.php.

Mauricio Orbes-Arteaga, M. Jorge Cardoso, Lauge Sørensen et al.

Segmenting vascular pathologies such as white matter lesions in Brain magnetic resonance images (MRIs) require acquisition of multiple sequences such as T1-weighted (T1-w) --on which lesions appear hypointense-- and fluid attenuated inversion recovery (FLAIR) sequence --where lesions appear hyperintense--. However, most of the existing retrospective datasets do not consist of FLAIR sequences. Existing missing modality imputation methods separate the process of imputation, and the process of segmentation. In this paper, we propose a method to link both modality imputation and segmentation using convolutional neural networks. We show that by jointly optimizing the imputation network and the segmentation network, the method not only produces more realistic synthetic FLAIR images from T1-w images, but also improves the segmentation of WMH from T1-w images only.

Mostafa Mehdipour Ghazi, Mads Nielsen, Akshay Pai et al.

Disease progression modeling (DPM) using longitudinal data is a challenging task in machine learning for healthcare that can provide clinicians with better tools for diagnosis and monitoring of disease. Existing DPM algorithms neglect temporal dependencies among measurements and make parametric assumptions about biomarker trajectories. In addition, they do not model multiple biomarkers jointly and need to align subjects' trajectories. In this paper, recurrent neural networks (RNNs) are utilized to address these issues. However, in many cases, longitudinal cohorts contain incomplete data, which hinders the application of standard RNNs and requires a pre-processing step such as imputation of the missing values. We, therefore, propose a generalized training rule for the most widely used RNN architecture, long short-term memory (LSTM) networks, that can handle missing values in both target and predictor variables. This algorithm is applied for modeling the progression of Alzheimer's disease (AD) using magnetic resonance imaging (MRI) biomarkers. The results show that the proposed LSTM algorithm achieves a lower mean absolute error for prediction of measurements across all considered MRI biomarkers compared to using standard LSTM networks with data imputation or using a regression-based DPM method. Moreover, applying linear discriminant analysis to the biomarkers' values predicted by the proposed algorithm results in a larger area under the receiver operating characteristic curve (AUC) for clinical diagnosis of AD compared to the same alternatives, and the AUC is comparable to state-of-the-art AUCs from a recent cross-sectional medical image classification challenge. This paper shows that built-in handling of missing values in LSTM network training paves the way for application of RNNs in disease progression modeling.

Akshay Pai, Stefan Sommer, Lars Lau Raket et al.

Template estimation plays a crucial role in computational anatomy since it provides reference frames for performing statistical analysis of the underlying anatomical population variability. While building models for template estimation, variability in sites and image acquisition protocols need to be accounted for. To account for such variability, we propose a generative template estimation model that makes simultaneous inference of both bias fields in individual images, deformations for image registration, and variance hyperparameters. In contrast, existing maximum a posterori based methods need to rely on either bias-invariant similarity measures or robust image normalization. Results on synthetic and real brain MRI images demonstrate the capability of the model to capture heterogeneity in intensities and provide a reliable template estimation from registration.

Malte Stær Nissen, Oswin Krause, Kristian Almstrup et al.

We compare a set of convolutional neural network (CNN) architectures for the task of segmenting and detecting human sperm cells in an image taken from a semen sample. In contrast to previous work, samples are not stained or washed to allow for full sperm quality analysis, making analysis harder due to clutter. Our results indicate that training on full images is superior to training on patches when class-skew is properly handled. Full image training including up-sampling during training proves to be beneficial in deep CNNs for pixel wise accuracy and detection performance. Predicted sperm cells are found by using connected components on the CNN predictions. We investigate optimization of a threshold parameter on the size of detected components. Our best network achieves 93.87% precision and 91.89% recall on our test dataset after thresholding outperforming a classical mage analysis approach.

Aasa Feragen, Pechin Lo, Marleen de Bruijne et al.

In order to develop statistical methods for shapes with a tree-structure, we construct a shape space framework for tree-like shapes and study metrics on the shape space. This shape space has singularities, corresponding to topological transitions in the represented trees. We study two closely related metrics on the shape space, TED and QED. QED is a quotient Euclidean distance arising naturally from the shape space formulation, while TED is the classical tree edit distance. Using Gromov's metric geometry we gain new insight into the geometries defined by TED and QED. We show that the new metric QED has nice geometric properties which facilitate statistical analysis, such as existence and local uniqueness of geodesics and averages. TED, on the other hand, does not share the geometric advantages of QED, but has nice algorithmic properties. We provide a theoretical framework and experimental results on synthetic data trees as well as airway trees from pulmonary CT scans. This way, we effectively illustrate that our framework has both the theoretical and qualitative properties necessary to build a theory of statistical tree-shape analysis.