Florian Dubost

Ali Mirzazadeh, Florian Dubost, Maxwell Pike et al. · stanford

Attention--or attribution--maps methods are methods designed to highlight regions of the model's input that were discriminative for its predictions. However, different attention maps methods can highlight different regions of the input, with sometimes contradictory explanations for a prediction. This effect is exacerbated when the training set is small. This indicates that either the model learned incorrect representations or that the attention maps methods did not accurately estimate the model's representations. We propose an unsupervised fine-tuning method that optimizes the consistency of attention maps and show that it improves both classification performance and the quality of attention maps. We propose an implementation for two state-of-the-art attention computation methods, Grad-CAM and Guided Backpropagation, which relies on an input masking technique. We also show results on Grad-CAM and Integrated Gradients in an ablation study. We evaluate this method on our own dataset of event detection in continuous video recordings of hospital patients aggregated and curated for this work. As a sanity check, we also evaluate the proposed method on PASCAL VOC and SVHN. With the proposed method, with small training sets, we achieve a 6.6 points lift of F1 score over the baselines on our video dataset, a 2.9 point lift of F1 score on PASCAL, and a 1.8 points lift of mean Intersection over Union over Grad-CAM for weakly supervised detection on PASCAL. Those improved attention maps may help clinicians better understand vision model predictions and ease the deployment of machine learning systems into clinical care. We share part of the code for this article at the following repository: https://github.com/alimirzazadeh/SemisupervisedAttention.

Jianning Li, Zongwei Zhou, Jiancheng Yang et al.

Prior to the deep learning era, shape was commonly used to describe the objects. Nowadays, state-of-the-art (SOTA) algorithms in medical imaging are predominantly diverging from computer vision, where voxel grids, meshes, point clouds, and implicit surface models are used. This is seen from numerous shape-related publications in premier vision conferences as well as the growing popularity of ShapeNet (about 51,300 models) and Princeton ModelNet (127,915 models). For the medical domain, we present a large collection of anatomical shapes (e.g., bones, organs, vessels) and 3D models of surgical instrument, called MedShapeNet, created to facilitate the translation of data-driven vision algorithms to medical applications and to adapt SOTA vision algorithms to medical problems. As a unique feature, we directly model the majority of shapes on the imaging data of real patients. As of today, MedShapeNet includes 23 dataset with more than 100,000 shapes that are paired with annotations (ground truth). Our data is freely accessible via a web interface and a Python application programming interface (API) and can be used for discriminative, reconstructive, and variational benchmarks as well as various applications in virtual, augmented, or mixed reality, and 3D printing. Exemplary, we present use cases in the fields of classification of brain tumors, facial and skull reconstructions, multi-class anatomy completion, education, and 3D printing. In future, we will extend the data and improve the interfaces. The project pages are: https://medshapenet.ikim.nrw/ and https://github.com/Jianningli/medshapenet-feedback

Soumick Chatterjee, Hadya Yassin, Florian Dubost et al.

Deep learning has demonstrated significant potential in medical imaging; however, the opacity of "black-box" models hinders clinical trust, while segmentation tasks typically necessitate labourious, hard-to-obtain pixel-wise annotations. To address these challenges simultaneously, this paper introduces a framework for three inherently explainable classifiers (GP-UNet, GP-ShuffleUNet, and GP-ReconResNet). By integrating a global pooling mechanism, these networks generate localisation heatmaps that directly influence classification decisions, offering inherent interpretability without relying on potentially unreliable post-hoc methods. These heatmaps are subsequently thresholded to achieve weakly-supervised segmentation, requiring only image-level classification labels for training. Validated on two datasets for multi-class brain tumour classification, the proposed models achieved a peak F1-score of 0.93. For the weakly-supervised segmentation task, a median Dice score of 0.728 (95% CI 0.715-0.739) was recorded. Notably, on a subset of tumour-only images, the best model achieved an accuracy of 98.7%, outperforming state-of-the-art glioma grading binary classifiers. Furthermore, comparative Precision-Recall analysis validated the framework's robustness against severe class imbalance, establishing a direct correlation between diagnostic confidence and segmentation fidelity. These results demonstrate that the proposed framework successfully combines high diagnostic accuracy with essential transparency, offering a promising direction for trustworthy clinical decision support. Code is available on GitHub: https://github.com/soumickmj/GPModels

Carole H. Sudre, Kimberlin Van Wijnen, Florian Dubost et al.

Imaging markers of cerebral small vessel disease provide valuable information on brain health, but their manual assessment is time-consuming and hampered by substantial intra- and interrater variability. Automated rating may benefit biomedical research, as well as clinical assessment, but diagnostic reliability of existing algorithms is unknown. Here, we present the results of the \textit{VAscular Lesions DetectiOn and Segmentation} (\textit{Where is VALDO?}) challenge that was run as a satellite event at the international conference on Medical Image Computing and Computer Aided Intervention (MICCAI) 2021. This challenge aimed to promote the development of methods for automated detection and segmentation of small and sparse imaging markers of cerebral small vessel disease, namely enlarged perivascular spaces (EPVS) (Task 1), cerebral microbleeds (Task 2) and lacunes of presumed vascular origin (Task 3) while leveraging weak and noisy labels. Overall, 12 teams participated in the challenge proposing solutions for one or more tasks (4 for Task 1 - EPVS, 9 for Task 2 - Microbleeds and 6 for Task 3 - Lacunes). Multi-cohort data was used in both training and evaluation. Results showed a large variability in performance both across teams and across tasks, with promising results notably for Task 1 - EPVS and Task 2 - Microbleeds and not practically useful results yet for Task 3 - Lacunes. It also highlighted the performance inconsistency across cases that may deter use at an individual level, while still proving useful at a population level.

Erfan Darzi, Florian Dubost, N. M. Sijtsema et al.

Federated learning offers a privacy-preserving framework for medical image analysis but exposes the system to adversarial attacks. This paper aims to evaluate the vulnerabilities of federated learning networks in medical image analysis against such attacks. Employing domain-specific MRI tumor and pathology imaging datasets, we assess the effectiveness of known threat scenarios in a federated learning environment. Our tests reveal that domain-specific configurations can increase the attacker's success rate significantly. The findings emphasize the urgent need for effective defense mechanisms and suggest a critical re-evaluation of current security protocols in federated medical image analysis systems.

Erfan Darzi, Florian Dubost, Nanna. M. Sijtsema et al.

In this paper, we delve into the susceptibility of federated medical image analysis systems to adversarial attacks. Our analysis uncovers a novel exploitation avenue: using gradient information from prior global model updates, adversaries can enhance the efficiency and transferability of their attacks. Specifically, we demonstrate that single-step attacks (e.g. FGSM), when aptly initialized, can outperform the efficiency of their iterative counterparts but with reduced computational demand. Our findings underscore the need to revisit our understanding of AI security in federated healthcare settings.

Junze Liu, Kun Qian, Florian Dubost et al.

Vision-language models (VLMs) exhibit a striking paradox: they can generate executable code that reconstructs a 3D scene from geometric primitives with correct object counts, classes, and approximate positions, yet the same models fail at simpler spatial questions on the same image. We show that 3D geometric primitives (cubes, spheres, cylinders, expressed in executable code) serve as a powerful intermediate representation for spatial understanding, and exploit this through three contributions. First, we introduce \textbf{\textsc{SpatialBabel}}, a benchmark evaluating fourteen VLMs on primitive-based 3D scene reconstruction across six \emph{scene-code languages} (programming languages and declarative formats for 3D primitive scenes), revealing that a single model's object-detection F1 can vary by up to $5.7\times$ across languages. Second, we propose \textbf{Code-CoT} (Code Chain-of-Thought), a training-free inference strategy that routes spatial reasoning through primitive-based code generation. Code-CoT lifts the SpatialBabel-QA-Score by up to $+6.4$\% on primitive scenes and real-photo CV-Bench-3D accuracy by $+5.0$\% for VLMs with strong coding capabilities. Third, we propose \textbf{S$^{3}$-FT} (Self-Supervised Spatial Fine-Tuning), which self-supervisedly distills primitive spatial knowledge into general visual reasoning by parsing the model's own Three.js primitive-reconstructions into structured annotations and fine-tuning on the result, with \emph{no human labels and no teacher model}. Training on primitive images alone, S$^3$-FT improves Qwen3-VL-8B by $+4.6$ to $+8.6$\% on SpatialBabel-Primitive-QA, $+9.7$\% on CV-Bench-2D, and $+17$\% on HallusionBench; the recipe transfers across model families. These results establish geometric primitives in code as both a diagnostic and a transferable spatial vocabulary for VLMs. We will release all artifacts upon publication.

Yang Zheng, Qingqing Zhao, Guandao Yang et al.

Modeling and rendering photorealistic avatars is of crucial importance in many applications. Existing methods that build a 3D avatar from visual observations, however, struggle to reconstruct clothed humans. We introduce PhysAvatar, a novel framework that combines inverse rendering with inverse physics to automatically estimate the shape and appearance of a human from multi-view video data along with the physical parameters of the fabric of their clothes. For this purpose, we adopt a mesh-aligned 4D Gaussian technique for spatio-temporal mesh tracking as well as a physically based inverse renderer to estimate the intrinsic material properties. PhysAvatar integrates a physics simulator to estimate the physical parameters of the garments using gradient-based optimization in a principled manner. These novel capabilities enable PhysAvatar to create high-quality novel-view renderings of avatars dressed in loose-fitting clothes under motions and lighting conditions not seen in the training data. This marks a significant advancement towards modeling photorealistic digital humans using physically based inverse rendering with physics in the loop. Our project website is at: https://qingqing-zhao.github.io/PhysAvatar

Soumick Chatterjee, Hendrik Mattern, Marc Dörner et al.

The human brain receives nutrients and oxygen through an intricate network of blood vessels. Pathology affecting small vessels, at the mesoscopic scale, represents a critical vulnerability within the cerebral blood supply and can lead to severe conditions, such as Cerebral Small Vessel Diseases. The advent of 7 Tesla MRI systems has enabled the acquisition of higher spatial resolution images, making it possible to visualise such vessels in the brain. However, the lack of publicly available annotated datasets has impeded the development of robust, machine learning-driven segmentation algorithms. To address this, the SMILE-UHURA challenge was organised. This challenge, held in conjunction with the ISBI 2023, in Cartagena de Indias, Colombia, aimed to provide a platform for researchers working on related topics. The SMILE-UHURA challenge addresses the gap in publicly available annotated datasets by providing an annotated dataset of Time-of-Flight angiography acquired with 7T MRI. This dataset was created through a combination of automated pre-segmentation and extensive manual refinement. In this manuscript, sixteen submitted methods and two baseline methods are compared both quantitatively and qualitatively on two different datasets: held-out test MRAs from the same dataset as the training data (with labels kept secret) and a separate 7T ToF MRA dataset where both input volumes and labels are kept secret. The results demonstrate that most of the submitted deep learning methods, trained on the provided training dataset, achieved reliable segmentation performance. Dice scores reached up to 0.838 $\pm$ 0.066 and 0.716 $\pm$ 0.125 on the respective datasets, with an average performance of up to 0.804 $\pm$ 0.15.

Colton Stearns, Adam Harley, Mikaela Uy et al.

Gaussian splatting has become a popular representation for novel-view synthesis, exhibiting clear strengths in efficiency, photometric quality, and compositional edibility. Following its success, many works have extended Gaussians to 4D, showing that dynamic Gaussians maintain these benefits while also tracking scene geometry far better than alternative representations. Yet, these methods assume dense multi-view videos as supervision. In this work, we are interested in extending the capability of Gaussian scene representations to casually captured monocular videos. We show that existing 4D Gaussian methods dramatically fail in this setup because the monocular setting is underconstrained. Building off this finding, we propose a method we call Dynamic Gaussian Marbles, which consist of three core modifications that target the difficulties of the monocular setting. First, we use isotropic Gaussian "marbles'', reducing the degrees of freedom of each Gaussian. Second, we employ a hierarchical divide and-conquer learning strategy to efficiently guide the optimization towards solutions with globally coherent motion. Finally, we add image-level and geometry-level priors into the optimization, including a tracking loss that takes advantage of recent progress in point tracking. By constraining the optimization, Dynamic Gaussian Marbles learns Gaussian trajectories that enable novel-view rendering and accurately capture the 3D motion of the scene elements. We evaluate on the Nvidia Dynamic Scenes dataset and the DyCheck iPhone dataset, and show that Gaussian Marbles significantly outperforms other Gaussian baselines in quality, and is on-par with non-Gaussian representations, all while maintaining the efficiency, compositionality, editability, and tracking benefits of Gaussians. Our project page can be found here https://geometry.stanford.edu/projects/dynamic-gaussian-marbles.github.io/.

Siddharth Sharma, Florian Dubost, Christopher Lee-Messer et al.

In a clinical setting, epilepsy patients are monitored via video electroencephalogram (EEG) tests. A video EEG records what the patient experiences on videotape while an EEG device records their brainwaves. Currently, there are no existing automated methods for tracking the patient's location during a seizure, and video recordings of hospital patients are substantially different from publicly available video benchmark datasets. For example, the camera angle can be unusual, and patients can be partially covered with bedding sheets and electrode sets. Being able to track a patient in real-time with video EEG would be a promising innovation towards improving the quality of healthcare. Specifically, an automated patient detection system could supplement clinical oversight and reduce the resource-intensive efforts of nurses and doctors who need to continuously monitor patients. We evaluate an ImageNet pre-trained Mask R-CNN, a standard deep learning model for object detection, on the task of patient detection using our own curated dataset of 45 videos of hospital patients. The dataset was aggregated and curated for this work. We show that without fine-tuning, ImageNet pre-trained Mask R-CNN models perform poorly on such data. By fine-tuning the models with a subset of our dataset, we observe a substantial improvement in patient detection performance, with a mean average precision of 0.64. We show that the results vary substantially depending on the video clip.

Gerda Bortsova, Daniel Bos, Florian Dubost et al.

Purpose: To evaluate a fully-automated deep-learning-based method for assessment of intracranial carotid artery calcification (ICAC). Methods: Two observers manually delineated ICAC in non-contrast CT scans of 2,319 participants (mean age 69 (SD 7) years; 1154 women) of the Rotterdam Study, prospectively collected between 2003 and 2006. These data were used to retrospectively develop and validate a deep-learning-based method for automated ICAC delineation and volume measurement. To evaluate the method, we compared manual and automatic assessment (computed using ten-fold cross-validation) with respect to 1) the agreement with an independent observer's assessment (available in a random subset of 47 scans); 2) the accuracy in delineating ICAC as judged via blinded visual comparison by an expert; 3) the association with first stroke incidence from the scan date until 2012. All method performance metrics were computed using 10-fold cross-validation. Results: The automated delineation of ICAC reached sensitivity of 83.8% and positive predictive value (PPV) of 88%. The intraclass correlation between automatic and manual ICAC volume measures was 0.98 (95% CI: 0.97, 0.98; computed in the entire dataset). Measured between the assessments of independent observers, sensitivity was 73.9%, PPV was 89.5%, and intraclass correlation was 0.91 (95% CI: 0.84, 0.95; computed in the 47-scan subset). In the blinded visual comparisons, automatic delineations were more accurate than manual ones (p-value = 0.01). The association of ICAC volume with incident stroke was similarly strong for both automated (hazard ratio, 1.38 (95% CI: 1.12, 1.75) and manually measured volumes (hazard ratio, 1.48 (95% CI: 1.20, 1.87)). Conclusions: The developed model was capable of automated segmentation and volume quantification of ICAC with accuracy comparable to human experts.

Florian Dubost, Erin Hong, Max Pike et al.

Optimization plays a key role in the training of deep neural networks. Deciding when to stop training can have a substantial impact on the performance of the network during inference. Under certain conditions, the generalization error can display a double descent pattern during training: the learning curve is non-monotonic and seemingly diverges before converging again after additional epochs. This optimization pattern can lead to early stopping procedures to stop training before the second convergence and consequently select a suboptimal set of parameters for the network, with worse performance during inference. In this work, in addition to confirming that double descent occurs with small datasets and noisy labels as evidenced by others, we show that noisy labels must be present both in the training and generalization sets to observe a double descent pattern. We also show that the learning rate has an influence on double descent, and study how different optimizers and optimizer parameters influence the apparition of double descent. Finally, we show that increasing the learning rate can create an aliasing effect that masks the double descent pattern without suppressing it. We study this phenomenon through extensive experiments on variants of CIFAR-10 and show that they translate to a real world application: the forecast of seizure events in epileptic patients from continuous electroencephalographic recordings.

Siyi Tang, Jared A. Dunnmon, Khaled Saab et al.

Automated seizure detection and classification from electroencephalography (EEG) can greatly improve seizure diagnosis and treatment. However, several modeling challenges remain unaddressed in prior automated seizure detection and classification studies: (1) representing non-Euclidean data structure in EEGs, (2) accurately classifying rare seizure types, and (3) lacking a quantitative interpretability approach to measure model ability to localize seizures. In this study, we address these challenges by (1) representing the spatiotemporal dependencies in EEGs using a graph neural network (GNN) and proposing two EEG graph structures that capture the electrode geometry or dynamic brain connectivity, (2) proposing a self-supervised pre-training method that predicts preprocessed signals for the next time period to further improve model performance, particularly on rare seizure types, and (3) proposing a quantitative model interpretability approach to assess a model's ability to localize seizures within EEGs. When evaluating our approach on seizure detection and classification on a large public dataset, we find that our GNN with self-supervised pre-training achieves 0.875 Area Under the Receiver Operating Characteristic Curve on seizure detection and 0.749 weighted F1-score on seizure classification, outperforming previous methods for both seizure detection and classification. Moreover, our self-supervised pre-training strategy significantly improves classification of rare seizure types. Furthermore, quantitative interpretability analysis shows that our GNN with self-supervised pre-training precisely localizes 25.4% focal seizures, a 21.9 point improvement over existing CNNs. Finally, by superimposing the identified seizure locations on both raw EEG signals and EEG graphs, our approach could provide clinicians with an intuitive visualization of localized seizure regions.

Gerda Bortsova, Florian Dubost, Laurens Hogeweg et al.

Adversarial attacks consist in maliciously changing the input data to mislead the predictions of automated decision systems and are potentially a serious threat for automated medical image analysis. Previous studies have shown that it is possible to adversarially manipulate automated segmentations produced by neural networks in a targeted manner in the white-box attack setting. In this article, we studied the effectiveness of adversarial attacks in targeted modification of segmentations of anatomical structures in chest X-rays. Firstly, we experimented with using anatomically implausible shapes as targets for adversarial manipulation. We showed that, by adding almost imperceptible noise to the image, we can reliably force state-of-the-art neural networks to segment the heart as a heart symbol instead of its real anatomical shape. Moreover, such heart-shaping attack did not appear to require higher adversarial noise level than an untargeted attack based the same attack method. Secondly, we attempted to explore the limits of adversarial manipulation of segmentations. For that, we assessed the effectiveness of shrinking and enlarging segmentation contours for the three anatomical structures. We observed that adversarially extending segmentations of structures into regions with intensity and texture uncharacteristic for them presented a challenge to our attacks, as well as, in some cases, changing segmentations in ways that conflict with class adjacency priors learned by the target network. Additionally, we evaluated performances of the untargeted attacks and targeted heart attacks in the black-box attack scenario, using a surrogate network trained on a different subset of images. In both cases, the attacks were substantially less effective. We believe these findings bring novel insights into the current capabilities and limits of adversarial attacks for semantic segmentation.

Florian Dubost, Erin Hong, Nandita Bhaskhar et al.

Labeled data is a critical resource for training and evaluating machine learning models. However, many real-life datasets are only partially labeled. We propose a semi-supervised machine learning training strategy to improve event detection performance on sequential data, such as video recordings, when only sparse labels are available, such as event start times without their corresponding end times. Our method uses noisy guesses of the events' end times to train event detection models. Depending on how conservative these guesses are, mislabeled samples may be introduced into the training set. We further propose a mathematical model for explaining and estimating the evolution of the classification performance for increasingly noisier end time estimates. We show that neural networks can improve their detection performance by leveraging more training data with less conservative approximations despite the higher proportion of incorrect labels. We adapt sequential versions of CIFAR-10 and MNIST, and use the Berkeley MHAD and HMBD51 video datasets to empirically evaluate our method, and find that our risk-tolerant strategy outperforms conservative estimates by 3.5 points of mean average precision for CIFAR, 30 points for MNIST, 3 points for MHAD, and 14 points for HMBD51. Then, we leverage the proposed training strategy to tackle a real-life application: processing continuous video recordings of epilepsy patients, and show that our method outperforms baseline labeling methods by 17 points of average precision, and reaches a classification performance similar to that of fully supervised models. We share part of the code for this article.

Soumick Chatterjee, Kartik Prabhu, Mahantesh Pattadkal et al.

Blood vessels of the brain provide the human brain with the required nutrients and oxygen. As a vulnerable part of the cerebral blood supply, pathology of small vessels can cause serious problems such as Cerebral Small Vessel Diseases (CSVD). It has also been shown that CSVD is related to neurodegeneration, such as Alzheimer's disease. With the advancement of 7 Tesla MRI systems, higher spatial image resolution can be achieved, enabling the depiction of very small vessels in the brain. Non-Deep Learning-based approaches for vessel segmentation, e.g., Frangi's vessel enhancement with subsequent thresholding, are capable of segmenting medium to large vessels but often fail to segment small vessels. The sensitivity of these methods to small vessels can be increased by extensive parameter tuning or by manual corrections, albeit making them time-consuming, laborious, and not feasible for larger datasets. This paper proposes a deep learning architecture to automatically segment small vessels in 7 Tesla 3D Time-of-Flight (ToF) Magnetic Resonance Angiography (MRA) data. The algorithm was trained and evaluated on a small imperfect semi-automatically segmented dataset of only 11 subjects; using six for training, two for validation, and three for testing. The deep learning model based on U-Net Multi-Scale Supervision was trained using the training subset and was made equivariant to elastic deformations in a self-supervised manner using deformation-aware learning to improve the generalisation performance. The proposed technique was evaluated quantitatively and qualitatively against the test set and achieved a Dice score of 80.44 $\pm$ 0.83. Furthermore, the result of the proposed method was compared against a selected manually segmented region (62.07 resultant Dice) and has shown a considerable improvement (18.98\%) with deformation-aware learning.

Gerda Bortsova, Cristina González-Gonzalo, Suzanne C. Wetstein et al.

Adversarial attacks are considered a potentially serious security threat for machine learning systems. Medical image analysis (MedIA) systems have recently been argued to be vulnerable to adversarial attacks due to strong financial incentives and the associated technological infrastructure. In this paper, we study previously unexplored factors affecting adversarial attack vulnerability of deep learning MedIA systems in three medical domains: ophthalmology, radiology, and pathology. We focus on adversarial black-box settings, in which the attacker does not have full access to the target model and usually uses another model, commonly referred to as surrogate model, to craft adversarial examples. We consider this to be the most realistic scenario for MedIA systems. Firstly, we study the effect of weight initialization (ImageNet vs. random) on the transferability of adversarial attacks from the surrogate model to the target model. Secondly, we study the influence of differences in development data between target and surrogate models. We further study the interaction of weight initialization and data differences with differences in model architecture. All experiments were done with a perturbation degree tuned to ensure maximal transferability at minimal visual perceptibility of the attacks. Our experiments show that pre-training may dramatically increase the transferability of adversarial examples, even when the target and surrogate's architectures are different: the larger the performance gain using pre-training, the larger the transferability. Differences in the development data between target and surrogate models considerably decrease the performance of the attack; this decrease is further amplified by difference in the model architecture. We believe these factors should be considered when developing security-critical MedIA systems planned to be deployed in clinical practice.

Subhradeep Kayal, Florian Dubost, Harm A. W. M. Tiddens et al.

Data augmentation is of paramount importance in biomedical image processing tasks, characterized by inadequate amounts of labelled data, to best use all of the data that is present. In-use techniques range from intensity transformations and elastic deformations, to linearly combining existing data points to make new ones. In this work, we propose the use of spectral techniques for data augmentation, using the discrete cosine and wavelet transforms. We empirically evaluate our approaches on a CT texture analysis task to detect abnormal lung-tissue in patients with cystic fibrosis. Empirical experiments show that the proposed spectral methods perform favourably as compared to the existing methods. When used in combination with existing methods, our proposed approach can increase the relative minor class segmentation performance by 44.1% over a simple replication baseline.

Oliver Werner, Kimberlin M. H. van Wijnen, Wiro J. Niessen et al.

To measure the volume of specific image structures, a typical approach is to first segment those structures using a neural network trained on voxel-wise (strong) labels and subsequently compute the volume from the segmentation. A more straightforward approach would be to predict the volume directly using a neural network based regression approach, trained on image-level (weak) labels indicating volume. In this article, we compared networks optimized with weak and strong labels, and study their ability to generalize to other datasets. We experimented with white matter hyperintensity (WMH) volume prediction in brain MRI scans. Neural networks were trained on a large local dataset and their performance was evaluated on four independent public datasets. We showed that networks optimized using only weak labels reflecting WMH volume generalized better for WMH volume prediction than networks optimized with voxel-wise segmentations of WMH. The attention maps of networks trained with weak labels did not seem to delineate WMHs, but highlighted instead areas with smooth contours around or near WMHs. By correcting for possible confounders we showed that networks trained on weak labels may have learnt other meaningful features that are more suited to generalization to unseen data. Our results suggest that for imaging biomarkers that can be derived from segmentations, training networks to predict the biomarker directly may provide more robust results than solving an intermediate segmentation step.

Gerda Bortsova, Florian Dubost, Laurens Hogeweg et al.

The scarcity of labeled data often limits the application of supervised deep learning techniques for medical image segmentation. This has motivated the development of semi-supervised techniques that learn from a mixture of labeled and unlabeled images. In this paper, we propose a novel semi-supervised method that, in addition to supervised learning on labeled training images, learns to predict segmentations consistent under a given class of transformations on both labeled and unlabeled images. More specifically, in this work we explore learning equivariance to elastic deformations. We implement this through: 1) a Siamese architecture with two identical branches, each of which receives a differently transformed image, and 2) a composite loss function with a supervised segmentation loss term and an unsupervised term that encourages segmentation consistency between the predictions of the two branches. We evaluate the method on a public dataset of chest radiographs with segmentations of anatomical structures using 5-fold cross-validation. The proposed method reaches significantly higher segmentation accuracy compared to supervised learning. This is due to learning transformation consistency on both labeled and unlabeled images, with the latter contributing the most. We achieve the performance comparable to state-of-the-art chest X-ray segmentation methods while using substantially fewer labeled images.

Florian Dubost, Benjamin Collery, Antonin Renaudier et al.

Scoliosis is a condition defined by an abnormal spinal curvature. For diagnosis and treatment planning of scoliosis, spinal curvature can be estimated using Cobb angles. We propose an automated method for the estimation of Cobb angles from X-ray scans. First, the centerline of the spine was segmented using a cascade of two convolutional neural networks. After smoothing the centerline, Cobb angles were automatically estimated using the derivative of the centerline. We evaluated the results using the mean absolute error and the average symmetric mean absolute percentage error between the manual assessment by experts and the automated predictions. For optimization, we used 609 X-ray scans from the London Health Sciences Center, and for evaluation, we participated in the international challenge "Accurate Automated Spinal Curvature Estimation, MICCAI 2019" (100 scans). On the challenge's test set, we obtained an average symmetric mean absolute percentage error of 22.96.

Chaoping Zhang, Florian Dubost, Marleen de Bruijne et al.

Deep learning has been successfully demonstrated in MRI reconstruction of accelerated acquisitions. However, its dependence on representative training data limits the application across different contrasts, anatomies, or image sizes. To address this limitation, we propose an unsupervised, auto-calibrated k-space completion method, based on a uniquely designed neural network that reconstructs the full k-space from an undersampled k-space, exploiting the redundancy among the multiple channels in the receive coil in a parallel imaging acquisition. To achieve this, contrary to common convolutional network approaches, the proposed network has a decreasing number of feature maps of constant size. In contrast to conventional parallel imaging methods such as GRAPPA that estimate the prediction kernel from the fully sampled autocalibration signals in a linear way, our method is able to learn nonlinear relations between sampled and unsampled positions in k-space. The proposed method was compared to the start-of-the-art ESPIRiT and RAKI methods in terms of noise amplification and visual image quality in both phantom and in-vivo experiments. The experiments indicate that APIR-Net provides a promising alternative to the conventional parallel imaging methods, and results in improved image quality especially for low SNR acquisitions.

Kimberlin M. H. van Wijnen, Florian Dubost, Pinar Yilmaz et al.

Localization of focal vascular lesions on brain MRI is an important component of research on the etiology of neurological disorders. However, manual annotation of lesions can be challenging, time-consuming and subject to observer bias. Automated detection methods often need voxel-wise annotations for training. We propose a novel approach for automated lesion detection that can be trained on scans only annotated with a dot per lesion instead of a full segmentation. From the dot annotations and their corresponding intensity images we compute various distance maps (DMs), indicating the distance to a lesion based on spatial distance, intensity distance, or both. We train a fully convolutional neural network (FCN) to predict these DMs for unseen intensity images. The local optima in the predicted DMs are expected to correspond to lesion locations. We show the potential of this approach to detect enlarged perivascular spaces in white matter on a large brain MRI dataset with an independent test set of 1000 scans. Our method matches the intra-rater performance of the expert rater that was computed on an independent set. We compare the different types of distance maps, showing that incorporating intensity information in the distance maps used to train an FCN greatly improves performance.

Bernhard Egger, Markus D. Schirmer, Florian Dubost et al.

We propose and demonstrate a joint model of anatomical shapes, image features and clinical indicators for statistical shape modeling and medical image analysis. The key idea is to employ a copula model to separate the joint dependency structure from the marginal distributions of variables of interest. This separation provides flexibility on the assumptions made during the modeling process. The proposed method can handle binary, discrete, ordinal and continuous variables. We demonstrate a simple and efficient way to include binary, discrete and ordinal variables into the modeling. We build Bayesian conditional models based on observed partial clinical indicators, features or shape based on Gaussian processes capturing the dependency structure. We apply the proposed method on a stroke dataset to jointly model the shape of the lateral ventricles, the spatial distribution of the white matter hyperintensity associated with periventricular white matter disease, and clinical indicators. The proposed method yields interpretable joint models for data exploration and patient-specific statistical shape models for medical image analysis.

Florian Dubost, Marleen de Bruijne, Marco Nardin et al.

Registration is a core component of many imaging pipelines. In case of clinical scans, with lower resolution and sometimes substantial motion artifacts, registration can produce poor results. Visual assessment of registration quality in large clinical datasets is inefficient. In this work, we propose to automatically assess the quality of registration to an atlas in clinical FLAIR MRI scans of the brain. The method consists of automatically segmenting the ventricles of a given scan using a neural network, and comparing the segmentation to the atlas' ventricles propagated to image space. We used the proposed method to improve clinical image registration to a general atlas by computing multiple registrations and then selecting the registration that yielded the highest ventricle overlap. Methods were evaluated in a single-site dataset of more than 1000 scans, as well as a multi-center dataset comprising 142 clinical scans from 12 sites. The automated ventricle segmentation reached a Dice coefficient with manual annotations of 0.89 in the single-site dataset, and 0.83 in the multi-center dataset. Registration via age-specific atlases could improve ventricle overlap compared to a direct registration to the general atlas (Dice similarity coefficient increase up to 0.15). Experiments also showed that selecting scans with the registration quality assessment method could improve the quality of average maps of white matter hyperintensity burden, instead of using all scans for the computation of the white matter hyperintensity map. In this work, we demonstrated the utility of an automated tool for assessing image registration quality in clinical scans. This image quality assessment step could ultimately assist in the translation of automated neuroimaging pipelines to the clinic.

Florian Dubost, Hieab Adams, Pinar Yilmaz et al.

Finding automatically multiple lesions in large images is a common problem in medical image analysis. Solving this problem can be challenging if, during optimization, the automated method cannot access information about the location of the lesions nor is given single examples of the lesions. We propose a new weakly supervised detection method using neural networks, that computes attention maps revealing the locations of brain lesions. These attention maps are computed using the last feature maps of a segmentation network optimized only with global image-level labels. The proposed method can generate attention maps at full input resolution without need for interpolation during preprocessing, which allows small lesions to appear in attention maps. For comparison, we modify state-of-the-art methods to compute attention maps for weakly supervised object detection, by using a global regression objective instead of the more conventional classification objective. This regression objective optimizes the number of occurrences of the target object in an image, e.g. the number of brain lesions in a scan, or the number of digits in an image. We study the behavior of the proposed method in MNIST-based detection datasets, and evaluate it for the challenging detection of enlarged perivascular spaces - a type of brain lesion - in a dataset of 2202 3D scans with point-wise annotations in the center of all lesions in four brain regions. In the brain dataset, the weakly supervised detection methods come close to the human intrarater agreement in each region. The proposed method reaches the best area under the curve in two out of four regions, and has the lowest number of false positive detections in all regions, while its average sensitivity over all regions is similar to that of the other best methods. The proposed method can facilitate epidemiological and clinical studies of enlarged perivascular spaces.

Vikram Venkatraghavan, Florian Dubost, Esther E. Bron et al.

Event-based models (EBM) are a class of disease progression models that can be used to estimate temporal ordering of neuropathological changes from cross-sectional data. Current EBMs only handle scalar biomarkers, such as regional volumes, as inputs. However, regional aggregates are a crude summary of the underlying high-resolution images, potentially limiting the accuracy of EBM. Therefore, we propose a novel method that exploits high-dimensional voxel-wise imaging biomarkers: n-dimensional discriminative EBM (nDEBM). nDEBM is based on an insight that mixture modeling, which is a key element of conventional EBMs, can be replaced by a more scalable semi-supervised support vector machine (SVM) approach. This SVM is used to estimate the degree of abnormality of each region which is then used to obtain subject-specific disease progression patterns. These patterns are in turn used for estimating the mean ordering by fitting a generalized Mallows model. In order to validate the biomarker ordering obtained using nDEBM, we also present a framework for Simulation of Imaging Biomarkers' Temporal Evolution (SImBioTE) that mimics neurodegeneration in brain regions. SImBioTE trains variational auto-encoders (VAE) in different brain regions independently to simulate images at varying stages of disease progression. We also validate nDEBM clinically using data from the Alzheimer's Disease Neuroimaging Initiative (ADNI). In both experiments, nDEBM using high-dimensional features gave better performance than state-of-the-art EBM methods using regional volume biomarkers. This suggests that nDEBM is a promising approach for disease progression modeling.

Gerda Bortsova, Florian Dubost, Silas Ørting et al.

We propose an end-to-end deep learning method that learns to estimate emphysema extent from proportions of the diseased tissue. These proportions were visually estimated by experts using a standard grading system, in which grades correspond to intervals (label example: 1-5% of diseased tissue). The proposed architecture encodes the knowledge that the labels represent a volumetric proportion. A custom loss is designed to learn with intervals. Thus, during training, our network learns to segment the diseased tissue such that its proportions fit the ground truth intervals. Our architecture and loss combined improve the performance substantially (8% ICC) compared to a more conventional regression network. We outperform traditional lung densitometry and two recently published methods for emphysema quantification by a large margin (at least 7% AUC and 15% ICC), and achieve near-human-level performance. Moreover, our method generates emphysema segmentations that predict the spatial distribution of emphysema at human level.

Florian Dubost, Gerda Bortsova, Hieab Adams et al.

Deep learning techniques are often criticized to heavily depend on a large quantity of labeled data. This problem is even more challenging in medical image analysis where the annotator expertise is often scarce. We propose a novel data-augmentation method to regularize neural network regressors that learn from a single global label per image. The principle of the method is to create new samples by recombining existing ones. We demonstrate the performance of our algorithm on two tasks: estimation of the number of enlarged perivascular spaces in the basal ganglia, and estimation of white matter hyperintensities volume. We show that the proposed method improves the performance over more basic data augmentation. The proposed method reached an intraclass correlation coefficient between ground truth and network predictions of 0.73 on the first task and 0.84 on the second task, only using between 25 and 30 scans with a single global label per scan for training. With the same number of training scans, more conventional data augmentation methods could only reach intraclass correlation coefficients of 0.68 on the first task, and 0.79 on the second task.

Filipe Marques, Florian Dubost, Mariette Kemner-van de Corput et al.

Cystic fibrosis is a genetic disease which may appear in early life with structural abnormalities in lung tissues. We propose to detect these abnormalities using a texture classification approach. Our method is a cascade of two convolutional neural networks. The first network detects the presence of abnormal tissues. The second network identifies the type of the structural abnormalities: bronchiectasis, atelectasis or mucus plugging.We also propose a network computing pixel-wise heatmaps of abnormality presence learning only from the patch-wise annotations. Our database consists of CT scans of 194 subjects. We use 154 subjects to train our algorithms and the 40 remaining ones as a test set. We compare our method with random forest and a single neural network approach. The first network reaches an accuracy of 0,94 for disease detection, 0,18 higher than the random forest classifier and 0,37 higher than the single neural network. Our cascade approach yields a final class-averaged F1-score of 0,33, outperforming the baseline method and the single network by 0,10 and 0,12.

Florian Dubost, Hieab Adams, Gerda Bortsova et al.

Enlarged perivascular spaces (EPVS) in the brain are an emerging imaging marker for cerebral small vessel disease, and have been shown to be related to increased risk of various neurological diseases, including stroke and dementia. Automatic quantification of EPVS would greatly help to advance research into its etiology and its potential as a risk indicator of disease. We propose a convolutional network regression method to quantify the extent of EPVS in the basal ganglia from 3D brain MRI. We first segment the basal ganglia and subsequently apply a 3D convolutional regression network designed for small object detection within this region of interest. The network takes an image as input, and outputs a quantification score of EPVS. The network has significantly more convolution operations than pooling ones and no final activation, allowing it to span the space of real numbers. We validated our approach using a dataset of 2000 brain MRI scans scored visually. Experiments with varying sizes of training and test sets showed that a good performance can be achieved with a training set of only 200 scans. With a training set of 1000 scans, the intraclass correlation coefficient (ICC) between our scoring method and the expert's visual score was 0.74. Our method outperforms by a large margin - more than 0.10 - four more conventional automated approaches based on intensities, scale-invariant feature transform, and random forest. We show that the network learns the structures of interest and investigate the influence of hyper-parameters on the performance. We also evaluate the reproducibility of our network using a set of 60 subjects scanned twice (scan-rescan reproducibility). On this set our network achieves an ICC of 0.93, while the intrarater agreement reaches 0.80. Furthermore, the automatic EPVS scoring correlates similarly to age as visual scoring.

Gerda Bortsova, Gijs van Tulder, Florian Dubost et al.

Intracranial carotid artery calcification (ICAC) is a major risk factor for stroke, and might contribute to dementia and cognitive decline. Reliance on time-consuming manual annotation of ICAC hampers much demanded further research into the relationship between ICAC and neurological diseases. Automation of ICAC segmentation is therefore highly desirable, but difficult due to the proximity of the lesions to bony structures with a similar attenuation coefficient. In this paper, we propose a method for automatic segmentation of ICAC; the first to our knowledge. Our method is based on a 3D fully convolutional neural network that we extend with two regularization techniques. Firstly, we use deep supervision (hidden layers supervision) to encourage discriminative features in the hidden layers. Secondly, we augment the network with skip connections, as in the recently developed ResNet, and dropout layers, inserted in a way that skip connections circumvent them. We investigate the effect of skip connections and dropout. In addition, we propose a simple problem-specific modification of the network objective function that restricts the focus to the most important image regions and simplifies the optimization. We train and validate our model using 882 CT scans and test on 1,000. Our regularization techniques and objective improve the average Dice score by 7.1%, yielding an average Dice of 76.2% and 97.7% correlation between predicted ICAC volumes and manual annotations.

Florian Dubost, Gerda Bortsova, Hieab Adams et al.

We propose a novel convolutional neural network for lesion detection from weak labels. Only a single, global label per image - the lesion count - is needed for training. We train a regression network with a fully convolutional architecture combined with a global pooling layer to aggregate the 3D output into a scalar indicating the lesion count. When testing on unseen images, we first run the network to estimate the number of lesions. Then we remove the global pooling layer to compute localization maps of the size of the input image. We evaluate the proposed network on the detection of enlarged perivascular spaces in the basal ganglia in MRI. Our method achieves a sensitivity of 62% with on average 1.5 false positives per image. Compared with four other approaches based on intensity thresholding, saliency and class maps, our method has a 20% higher sensitivity.

Florian Dubost, Loic Peter, Christian Rupprecht et al.

We propose a novel hands-free method to interactively segment 3D medical volumes. In our scenario, a human user progressively segments an organ by answering a series of questions of the form "Is this voxel inside the object to segment?". At each iteration, the chosen question is defined as the one halving a set of candidate segmentations given the answered questions. For a quick and efficient exploration, these segmentations are sampled according to the Metropolis-Hastings algorithm. Our sampling technique relies on a combination of relaxed shape prior, learnt probability map and consistency with previous answers. We demonstrate the potential of our strategy on a prostate segmentation MRI dataset. Through the study of failure cases with synthetic examples, we demonstrate the adaptation potential of our method. We also show that our method outperforms two intuitive baselines: one based on random questions, the other one being the thresholded probability map.