Behzad Bozorgtabar

Thomas Stegmüller, Tim Lebailly, Behzad Bozorgtabar et al.

Learning dense visual representations without labels is an arduous task and more so from scene-centric data. We propose to tackle this challenging problem by proposing a Cross-view consistency objective with an Online Clustering mechanism (CrOC) to discover and segment the semantics of the views. In the absence of hand-crafted priors, the resulting method is more generalizable and does not require a cumbersome pre-processing step. More importantly, the clustering algorithm conjointly operates on the features of both views, thereby elegantly bypassing the issue of content not represented in both views and the ambiguous matching of objects from one crop to the other. We demonstrate excellent performance on linear and unsupervised segmentation transfer tasks on various datasets and similarly for video object segmentation. Our code and pre-trained models are publicly available at https://github.com/stegmuel/CrOC.

Thomas Stegmüller, Christian Abbet, Behzad Bozorgtabar et al.

Screening Papanicolaou test samples has proven to be highly effective in reducing cervical cancer-related mortality. However, the lack of trained cytopathologists hinders its widespread implementation in low-resource settings. Deep learning-based telecytology diagnosis emerges as an appealing alternative, but it requires the collection of large annotated training datasets, which is costly and time-consuming. In this paper, we demonstrate that the abundance of unlabeled images that can be extracted from Pap smear test whole slide images presents a fertile ground for self-supervised learning methods, yielding performance improvements relative to readily available pre-trained models for various downstream tasks. In particular, we propose \textbf{C}ervical \textbf{C}ell \textbf{C}opy-\textbf{P}asting ($\texttt{C}^{3}\texttt{P}$) as an effective augmentation method, which enables knowledge transfer from open-source and labeled single-cell datasets to unlabeled tiles. Not only does $\texttt{C}^{3}\texttt{P}$ outperforms naive transfer from single-cell images, but we also demonstrate its advantageous integration into multiple instance learning methods. Importantly, all our experiments are conducted on our introduced \textit{in-house} dataset comprising liquid-based cytology Pap smear images obtained using low-cost technologies. This aligns with our objective of leveraging deep learning-based telecytology for diagnosis in low-resource settings.

Tim Lebailly, Thomas Stegmüller, Behzad Bozorgtabar et al.

Leveraging nearest neighbor retrieval for self-supervised representation learning has proven beneficial with object-centric images. However, this approach faces limitations when applied to scene-centric datasets, where multiple objects within an image are only implicitly captured in the global representation. Such global bootstrapping can lead to undesirable entanglement of object representations. Furthermore, even object-centric datasets stand to benefit from a finer-grained bootstrapping approach. In response to these challenges, we introduce a novel Cross-Image Object-Level Bootstrapping method tailored to enhance dense visual representation learning. By employing object-level nearest neighbor bootstrapping throughout the training, CrIBo emerges as a notably strong and adequate candidate for in-context learning, leveraging nearest neighbor retrieval at test time. CrIBo shows state-of-the-art performance on the latter task while being highly competitive in more standard downstream segmentation tasks. Our code and pretrained models are publicly available at https://github.com/tileb1/CrIBo.

Tim Lebailly, Thomas Stegmüller, Behzad Bozorgtabar et al.

Most self-supervised methods for representation learning leverage a cross-view consistency objective i.e., they maximize the representation similarity of a given image's augmented views. Recent work NNCLR goes beyond the cross-view paradigm and uses positive pairs from different images obtained via nearest neighbor bootstrapping in a contrastive setting. We empirically show that as opposed to the contrastive learning setting which relies on negative samples, incorporating nearest neighbor bootstrapping in a self-distillation scheme can lead to a performance drop or even collapse. We scrutinize the reason for this unexpected behavior and provide a solution. We propose to adaptively bootstrap neighbors based on the estimated quality of the latent space. We report consistent improvements compared to the naive bootstrapping approach and the original baselines. Our approach leads to performance improvements for various self-distillation method/backbone combinations and standard downstream tasks. Our code is publicly available at https://github.com/tileb1/AdaSim.

Devavrat Tomar, Guillaume Vray, Behzad Bozorgtabar et al.

Most recent test-time adaptation methods focus on only classification tasks, use specialized network architectures, destroy model calibration or rely on lightweight information from the source domain. To tackle these issues, this paper proposes a novel Test-time Self-Learning method with automatic Adversarial augmentation dubbed TeSLA for adapting a pre-trained source model to the unlabeled streaming test data. In contrast to conventional self-learning methods based on cross-entropy, we introduce a new test-time loss function through an implicitly tight connection with the mutual information and online knowledge distillation. Furthermore, we propose a learnable efficient adversarial augmentation module that further enhances online knowledge distillation by simulating high entropy augmented images. Our method achieves state-of-the-art classification and segmentation results on several benchmarks and types of domain shifts, particularly on challenging measurement shifts of medical images. TeSLA also benefits from several desirable properties compared to competing methods in terms of calibration, uncertainty metrics, insensitivity to model architectures, and source training strategies, all supported by extensive ablations. Our code and models are available on GitHub.

Behzad Bozorgtabar, Dwarikanath Mahapatra, Jean-Philippe Thiran

Unsupervised anomaly detection in medical images such as chest radiographs is stepping into the spotlight as it mitigates the scarcity of the labor-intensive and costly expert annotation of anomaly data. However, nearly all existing methods are formulated as a one-class classification trained only on representations from the normal class and discard a potentially significant portion of the unlabeled data. This paper focuses on a more practical setting, dual distribution anomaly detection for chest X-rays, using the entire training data, including both normal and unlabeled images. Inspired by a modern self-supervised vision transformer model trained using partial image inputs to reconstruct missing image regions -- we propose AMAE, a two-stage algorithm for adaptation of the pre-trained masked autoencoder (MAE). Starting from MAE initialization, AMAE first creates synthetic anomalies from only normal training images and trains a lightweight classifier on frozen transformer features. Subsequently, we propose an adaptation strategy to leverage unlabeled images containing anomalies. The adaptation scheme is accomplished by assigning pseudo-labels to unlabeled images and using two separate MAE based modules to model the normative and anomalous distributions of pseudo-labeled images. The effectiveness of the proposed adaptation strategy is evaluated with different anomaly ratios in an unlabeled training set. AMAE leads to consistent performance gains over competing self-supervised and dual distribution anomaly detection methods, setting the new state-of-the-art on three public chest X-ray benchmarks: RSNA, NIH-CXR, and VinDr-CXR.

Pushpak Pati, Guillaume Jaume, Zeineb Ayadi et al.

The segmentation and automatic identification of histological regions of diagnostic interest offer a valuable aid to pathologists. However, segmentation methods are hampered by the difficulty of obtaining pixel-level annotations, which are tedious and expensive to obtain for Whole-Slide images (WSI). To remedy this, weakly supervised methods have been developed to exploit the annotations directly available at the image level. However, to our knowledge, none of these techniques is adapted to deal with WSIs. In this paper, we propose WholeSIGHT, a weakly-supervised method, to simultaneously segment and classify WSIs of arbitrary shapes and sizes. Formally, WholeSIGHT first constructs a tissue-graph representation of the WSI, where the nodes and edges depict tissue regions and their interactions, respectively. During training, a graph classification head classifies the WSI and produces node-level pseudo labels via post-hoc feature attribution. These pseudo labels are then used to train a node classification head for WSI segmentation. During testing, both heads simultaneously render class prediction and segmentation for an input WSI. We evaluated WholeSIGHT on three public prostate cancer WSI datasets. Our method achieved state-of-the-art weakly-supervised segmentation performance on all datasets while resulting in better or comparable classification with respect to state-of-the-art weakly-supervised WSI classification methods. Additionally, we quantify the generalization capability of our method in terms of segmentation and classification performance, uncertainty estimation, and model calibration.

Guillaume Vray, Devavrat Tomar, Jean-Philippe Thiran et al.

Developing computational pathology models is essential for reducing manual tissue typing from whole slide images, transferring knowledge from the source domain to an unlabeled, shifted target domain, and identifying unseen categories. We propose a practical setting by addressing the above-mentioned challenges in one fell swoop, i.e., source-free open-set domain adaptation. Our methodology focuses on adapting a pre-trained source model to an unlabeled target dataset and encompasses both closed-set and open-set classes. Beyond addressing the semantic shift of unknown classes, our framework also deals with a covariate shift, which manifests as variations in color appearance between source and target tissue samples. Our method hinges on distilling knowledge from a self-supervised vision transformer (ViT), drawing guidance from either robustly pre-trained transformer models or histopathology datasets, including those from the target domain. In pursuit of this, we introduce a novel style-based adversarial data augmentation, serving as hard positives for self-training a ViT, resulting in highly contextualized embeddings. Following this, we cluster semantically akin target images, with the source model offering weak pseudo-labels, albeit with uncertain confidence. To enhance this process, we present the closed-set affinity score (CSAS), aiming to correct the confidence levels of these pseudo-labels and to calculate weighted class prototypes within the contextualized embedding space. Our approach establishes itself as state-of-the-art across three public histopathological datasets for colorectal cancer assessment. Notably, our self-training method seamlessly integrates with open-set detection methods, resulting in enhanced performance in both closed-set and open-set recognition tasks.

Govinda Kolli, Adinath Madhavrao Dukre, Behzad Bozorgtabar et al.

Medical Vision-Language Models (VLMs) often hallucinate by generating responses based on language priors rather than visual evidence, posing risks in clinical applications. We propose Visual Grounding Score Guided Decoding (VGS-Decoding), a training-free method to mitigate hallucinations during inference. Our key insight is that hallucinated tokens maintain or increase their probability when visual information is degraded, while visually grounded tokens decrease in probability. We introduce the Visual Grounding Score (VGS), which measures each token's visual dependency by comparing distributions from original and distorted images. During decoding, we reweight probabilities by amplifying visually grounded tokens while suppressing hallucinations. Unlike fixed-weight contrastive methods, VGS-Decoding provides per-token adaptive control. Experiments on MIMIC-Diff-VQA and VQA-RAD across LLaVA-Med, CheXagent, and MedGemma demonstrate consistent improvements, with up to +9.12% overall gain and $+8.98\%$ in open-ended recall, while introducing only $2\times$ inference overhead and no additional training, making it practical for clinical deployment. Upon acceptance, code will be released publicly to facilitate reproducibility.

Behzad Bozorgtabar, Dwarikanath Mahapatra, Sudipta Roy et al.

Medical vision-language models (VLMs) are strong zero-shot recognizers for medical imaging, but their reliability under domain shift hinges on calibrated uncertainty with guarantees. Split conformal prediction (SCP) offers finite-sample coverage, yet prediction sets often become large (low efficiency) and class-wise coverage unbalanced-high class-conditioned coverage gap (CCV), especially in few-shot, imbalanced regimes; moreover, naively adapting to calibration labels breaks exchangeability and voids guarantees. We propose \texttt{\textbf{LATA}} (Laplacian-Assisted Transductive Adaptation), a \textit{training- and label-free} refinement that operates on the joint calibration and test pool by smoothing zero-shot probabilities over an image-image k-NN graph using a small number of CCCP mean-field updates, preserving SCP validity via a deterministic transform. We further introduce a \textit{failure-aware} conformal score that plugs into the vision-language uncertainty (ViLU) framework, providing instance-level difficulty and label plausibility to improve prediction set efficiency and class-wise balance at fixed coverage. \texttt{\textbf{LATA}} is black-box (no VLM updates), compute-light (windowed transduction, no backprop), and includes an optional prior knob that can run strictly label-free or, if desired, in a label-informed variant using calibration marginals once. Across \textbf{three} medical VLMs and \textbf{nine} downstream tasks, \texttt{\textbf{LATA}} consistently reduces set size and CCV while matching or tightening target coverage, outperforming prior transductive baselines and narrowing the gap to label-using methods, while using far less compute. Comprehensive ablations and qualitative analyses show that \texttt{\textbf{LATA}} sharpens zero-shot predictions without compromising exchangeability.

Kangxian Xie, Jiancheng Yang, Nandor Pinter et al.

Accurate 3D anatomical segmentation is essential for clinical diagnosis and surgical planning. However, automated models frequently generate suboptimal shape predictions due to factors such as limited and imbalanced training data, inadequate labeling quality, and distribution shifts between training and deployment settings. A natural solution is to iteratively refine the predicted shape based on the radiologists' verbal instructions. However, this is hindered by the scarcity of paired data that explicitly links erroneous shapes to corresponding corrective instructions. As an initial step toward addressing this limitation, we introduce CoWTalk, a benchmark comprising 3D arterial anatomies with controllable synthesized anatomical errors and their corresponding repairing instructions. Building on this benchmark, we further propose an iterative refinement model that represents 3D shapes as vector sets and interacts with textual instructions to progressively update the target shape. Experimental results demonstrate that our method achieves significant improvements over corrupted inputs and competitive baselines, highlighting the feasibility of language-driven clinician-in-the-loop refinement for 3D medical shapes modeling.

Simon Ging, Philipp Arnold, Sebastian Walter et al.

Recent 3D CT vision-language models align volumes with reports via contrastive pretraining, but typically rely on limited public data and provide only coarse global supervision. We train a 3D CT vision-language model on 98k report-volume pairs (50k patients) collected at a single hospital, combined with public datasets, using SigLIP-style contrastive pretraining together with prompt-based disease supervision in the shared vision-text embedding space. On CT-RATE, our model achieves state-of-the-art text-to-image retrieval (R@10 31.5 vs. 22.2) and competitive disease classification (AUC 83.8 vs. 83.8), with consistent results on Rad-ChestCT (AUC 77.0 vs. 77.3). We further observe that radiologists routinely reference specific images within their reports (e.g., ``series X, image Y''), linking textual descriptions to precise axial locations. We automatically mine 262k such snippet-slice pairs and introduce the task of intra-scan snippet localization -- predicting the axial depth referred to by a text snippet -- reducing mean absolute error to 36.3 mm at 12 mm feature resolution, compared with 67.0 mm for the best baseline. Adding this localization objective leaves retrieval and classification broadly unchanged within confidence bounds, yielding a single unified model for retrieval, classification, and intra-scan grounding.

Chentao Li, Behzad Bozorgtabar, Yifang Ping et al.

Multiple instance learning (MIL) has been widely used for representing whole-slide pathology images. However, spatial, semantic, and decision entanglements among instances limit its representation and interpretability. To address these challenges, we propose a latent factor grouping-boosted cluster-reasoning instance disentangled learning framework for whole-slide image (WSI) interpretable representation in three phases. First, we introduce a novel positive semi-definite latent factor grouping that maps instances into a latent subspace, effectively mitigating spatial entanglement in MIL. To alleviate semantic entanglement, we employs instance probability counterfactual inference and optimization via cluster-reasoning instance disentangling. Finally, we employ a generalized linear weighted decision via instance effect re-weighting to address decision entanglement. Extensive experiments on multicentre datasets demonstrate that our model outperforms all state-of-the-art models. Moreover, it attains pathologist-aligned interpretability through disentangled representations and a transparent decision-making process.

Devavrat Tomar, Guillaume Vray, Dwarikanath Mahapatra et al.

In this paper, we address the challenge of few-shot classification in histopathology whole slide images (WSIs) by utilizing foundational vision-language models (VLMs) and slide-level prompt learning. Given the gigapixel scale of WSIs, conventional multiple instance learning (MIL) methods rely on aggregation functions to derive slide-level (bag-level) predictions from patch representations, which require extensive bag-level labels for training. In contrast, VLM-based approaches excel at aligning visual embeddings of patches with candidate class text prompts but lack essential pathological prior knowledge. Our method distinguishes itself by utilizing pathological prior knowledge from language models to identify crucial local tissue types (patches) for WSI classification, integrating this within a VLM-based MIL framework. Our approach effectively aligns patch images with tissue types, and we fine-tune our model via prompt learning using only a few labeled WSIs per category. Experimentation on real-world pathological WSI datasets and ablation studies highlight our method's superior performance over existing MIL- and VLM-based methods in few-shot WSI classification tasks. Our code is publicly available at https://github.com/LTS5/SLIP.

Guillaume Vray, Devavrat Tomar, Xufeng Gao et al.

This paper introduces ReservoirTTA, a novel plug-in framework designed for prolonged test-time adaptation (TTA) in scenarios where the test domain continuously shifts over time, including cases where domains recur or evolve gradually. At its core, ReservoirTTA maintains a reservoir of domain-specialized models -- an adaptive test-time model ensemble -- that both detects new domains via online clustering over style features of incoming samples and routes each sample to the appropriate specialized model, and thereby enables domain-specific adaptation. This multi-model strategy overcomes key limitations of single model adaptation, such as catastrophic forgetting, inter-domain interference, and error accumulation, ensuring robust and stable performance on sustained non-stationary test distributions. Our theoretical analysis reveals key components that bound parameter variance and prevent model collapse, while our plug-in TTA module mitigates catastrophic forgetting of previously encountered domains. Extensive experiments on scene-level corruption benchmarks (ImageNet-C, CIFAR-10/100-C), object-level style shifts (DomainNet-126, PACS), and semantic segmentation (Cityscapes->ACDC) covering recurring and continuously evolving domain shifts -- show that ReservoirTTA substantially improves adaptation accuracy and maintains stable performance across prolonged, recurring shifts, outperforming state-of-the-art methods. Our code is publicly available at https://github.com/LTS5/ReservoirTTA.

Devavrat Tomar, Guillaume Vray, Jean-Philippe Thiran et al.

Recent test-time adaptation methods heavily rely on nuanced adjustments of batch normalization (BN) parameters. However, one critical assumption often goes overlooked: that of independently and identically distributed (i.i.d.) test batches with respect to unknown labels. This oversight leads to skewed BN statistics and undermines the reliability of the model under non-i.i.d. scenarios. To tackle this challenge, this paper presents a novel method termed 'Un-Mixing Test-Time Normalization Statistics' (UnMix-TNS). Our method re-calibrates the statistics for each instance within a test batch by mixing it with multiple distinct statistics components, thus inherently simulating the i.i.d. scenario. The core of this method hinges on a distinctive online unmixing procedure that continuously updates these statistics components by incorporating the most similar instances from new test batches. Remarkably generic in its design, UnMix-TNS seamlessly integrates with a wide range of leading test-time adaptation methods and pre-trained architectures equipped with BN layers. Empirical evaluations corroborate the robustness of UnMix-TNS under varied scenarios-ranging from single to continual and mixed domain shifts, particularly excelling with temporally correlated test data and corrupted non-i.i.d. real-world streams. This adaptability is maintained even with very small batch sizes or single instances. Our results highlight UnMix-TNS's capacity to markedly enhance stability and performance across various benchmarks. Our code is publicly available at https://github.com/devavratTomar/unmixtns.

Devavrat Tomar, Behzad Bozorgtabar, Manana Lortkipanidze et al.

In medical image segmentation, supervised deep networks' success comes at the cost of requiring abundant labeled data. While asking domain experts to annotate only one or a few of the cohort's images is feasible, annotating all available images is impractical. This issue is further exacerbated when pre-trained deep networks are exposed to a new image dataset from an unfamiliar distribution. Using available open-source data for ad-hoc transfer learning or hand-tuned techniques for data augmentation only provides suboptimal solutions. Motivated by atlas-based segmentation, we propose a novel volumetric self-supervised learning for data augmentation capable of synthesizing volumetric image-segmentation pairs via learning transformations from a single labeled atlas to the unlabeled data. Our work's central tenet benefits from a combined view of one-shot generative learning and the proposed self-supervised training strategy that cluster unlabeled volumetric images with similar styles together. Unlike previous methods, our method does not require input volumes at inference time to synthesize new images. Instead, it can generate diversified volumetric image-segmentation pairs from a prior distribution given a single or multi-site dataset. Augmented data generated by our method used to train the segmentation network provide significant improvements over state-of-the-art deep one-shot learning methods on the task of brain MRI segmentation. Ablation studies further exemplified that the proposed appearance model and joint training are crucial to synthesize realistic examples compared to existing medical registration methods. The code, data, and models are available at https://github.com/devavratTomar/SST.

Christian Abbet, Linda Studer, Andreas Fischer et al.

Supervised learning is constrained by the availability of labeled data, which are especially expensive to acquire in the field of digital pathology. Making use of open-source data for pre-training or using domain adaptation can be a way to overcome this issue. However, pre-trained networks often fail to generalize to new test domains that are not distributed identically due to tissue stainings, types, and textures variations. Additionally, current domain adaptation methods mainly rely on fully-labeled source datasets. In this work, we propose Self-Rule to Multi-Adapt (SRMA), which takes advantage of self-supervised learning to perform domain adaptation, and removes the necessity of fully-labeled source datasets. SRMA can effectively transfer the discriminative knowledge obtained from a few labeled source domain's data to a new target domain without requiring additional tissue annotations. Our method harnesses both domains' structures by capturing visual similarity with intra-domain and cross-domain self-supervision. Moreover, we present a generalized formulation of our approach that allows the framework to learn from multiple source domains. We show that our proposed method outperforms baselines for domain adaptation of colorectal tissue type classification \new{in single and multi-source settings}, and further validate our approach on an in-house clinical cohort. The code and trained models are available open-source: https://github.com/christianabbet/SRA.

Ankan Deria, Dwarikanath Mahapatra, Behzad Bozorgtabar et al.

Virtual try-on seeks to generate photorealistic images of individuals in desired garments, a task that must simultaneously preserve personal identity and garment fidelity for practical use in fashion retail and personalization. However, existing methods typically handle upper and lower garments separately, rely on heavy preprocessing, and often fail to preserve person-specific cues such as tattoos, accessories, and body shape-resulting in limited realism and flexibility. To this end, we introduce MuGa-VTON, a unified multi-garment diffusion framework that jointly models upper and lower garments together with person identity in a shared latent space. Specifically, we proposed three key modules: the Garment Representation Module (GRM) for capturing both garment semantics, the Person Representation Module (PRM) for encoding identity and pose cues, and the A-DiT fusion module, which integrates garment, person, and text-prompt features through a diffusion transformer. This architecture supports prompt-based customization, allowing fine-grained garment modifications with minimal user input. Extensive experiments on the VITON-HD and DressCode benchmarks demonstrate that MuGa-VTON outperforms existing methods in both qualitative and quantitative evaluations, producing high-fidelity, identity-preserving results suitable for real-world virtual try-on applications.

Thomas Stegmüller, Tim Lebailly, Nikola Dukic et al.

Zero-shot classification capabilities naturally arise in models trained within a vision-language contrastive framework. Despite their classification prowess, these models struggle in dense tasks like zero-shot open-vocabulary segmentation. This deficiency is often attributed to the absence of localization cues in captions and the intertwined nature of the learning process, which encompasses both image representation learning and cross-modality alignment. To tackle these issues, we propose SimZSS, a Simple framework for open-vocabulary Zero-Shot Segmentation. The method is founded on two key principles: i) leveraging frozen vision-only models that exhibit spatial awareness while exclusively aligning the text encoder and ii) exploiting the discrete nature of text and linguistic knowledge to pinpoint local concepts within captions. By capitalizing on the quality of the visual representations, our method requires only image-caption pairs datasets and adapts to both small curated and large-scale noisy datasets. When trained on COCO Captions across 8 GPUs, SimZSS achieves state-of-the-art results on 7 out of 8 benchmark datasets in less than 15 minutes.

Thomas Stegmüller, Behzad Bozorgtabar, Antoine Spahr et al.

Progress in digital pathology is hindered by high-resolution images and the prohibitive cost of exhaustive localized annotations. The commonly used paradigm to categorize pathology images is patch-based processing, which often incorporates multiple instance learning (MIL) to aggregate local patch-level representations yielding image-level prediction. Nonetheless, diagnostically relevant regions may only take a small fraction of the whole tissue, and current MIL-based approaches often process images uniformly, discarding the inter-patches interactions. To alleviate these issues, we propose ScoreNet, a new efficient transformer that exploits a differentiable recommendation stage to extract discriminative image regions and dedicate computational resources accordingly. The proposed transformer leverages the local and global attention of a few dynamically recommended high-resolution regions at an efficient computational cost. We further introduce a novel mixing data-augmentation, namely ScoreMix, by leveraging the image's semantic distribution to guide the data mixing and produce coherent sample-label pairs. ScoreMix is embarrassingly simple and mitigates the pitfalls of previous augmentations, which assume a uniform semantic distribution and risk mislabeling the samples. Thorough experiments and ablation studies on three breast cancer histology datasets of Haematoxylin & Eosin (H&E) have validated the superiority of our approach over prior arts, including transformer-based models on tumour regions-of-interest (TRoIs) classification. ScoreNet equipped with proposed ScoreMix augmentation demonstrates better generalization capabilities and achieves new state-of-the-art (SOTA) results with only 50% of the data compared to other mixing augmentation variants. Finally, ScoreNet yields high efficacy and outperforms SOTA efficient transformers, namely TransPath and SwinTransformer.

Mohammad Saeed Rad, Thomas Yu, Behzad Bozorgtabar et al.

Existing reference (RF)-based super-resolution (SR) models try to improve perceptual quality in SR under the assumption of the availability of high-resolution RF images paired with low-resolution (LR) inputs at testing. As the RF images should be similar in terms of content, colors, contrast, etc. to the test image, this hinders the applicability in a real scenario. Other approaches to increase the perceptual quality of images, including perceptual loss and adversarial losses, tend to dramatically decrease fidelity to the ground-truth through significant decreases in PSNR/SSIM. Addressing both issues, we propose a simple yet universal approach to improve the perceptual quality of the HR prediction from a pre-trained SR network on a given LR input by further fine-tuning the SR network on a subset of images from the training dataset with similar patterns of activation as the initial HR prediction, with respect to the filters of a feature extractor. In particular, we show the effects of fine-tuning on these images in terms of the perceptual quality and PSNR/SSIM values. Contrary to perceptually driven approaches, we demonstrate that the fine-tuned network produces a HR prediction with both greater perceptual quality and minimal changes to the PSNR/SSIM with respect to the initial HR prediction. Further, we present novel numerical experiments concerning the filters of SR networks, where we show through filter correlation, that the filters of the fine-tuned network from our method are closer to "ideal" filters, than those of the baseline network or a network fine-tuned on random images.

Devavrat Tomar, Manana Lortkipanidze, Guillaume Vray et al.

Despite the successes of deep neural networks on many challenging vision tasks, they often fail to generalize to new test domains that are not distributed identically to the training data. The domain adaptation becomes more challenging for cross-modality medical data with a notable domain shift. Given that specific annotated imaging modalities may not be accessible nor complete. Our proposed solution is based on the cross-modality synthesis of medical images to reduce the costly annotation burden by radiologists and bridge the domain gap in radiological images. We present a novel approach for image-to-image translation in medical images, capable of supervised or unsupervised (unpaired image data) setups. Built upon adversarial training, we propose a learnable self-attentive spatial normalization of the deep convolutional generator network's intermediate activations. Unlike previous attention-based image-to-image translation approaches, which are either domain-specific or require distortion of the source domain's structures, we unearth the importance of the auxiliary semantic information to handle the geometric changes and preserve anatomical structures during image translation. We achieve superior results for cross-modality segmentation between unpaired MRI and CT data for multi-modality whole heart and multi-modal brain tumor MRI (T1/T2) datasets compared to the state-of-the-art methods. We also observe encouraging results in cross-modality conversion for paired MRI and CT images on a brain dataset. Furthermore, a detailed analysis of the cross-modality image translation, thorough ablation studies confirm our proposed method's efficacy.

Valentin Anklin, Pushpak Pati, Guillaume Jaume et al.

Segmenting histology images into diagnostically relevant regions is imperative to support timely and reliable decisions by pathologists. To this end, computer-aided techniques have been proposed to delineate relevant regions in scanned histology slides. However, the techniques necessitate task-specific large datasets of annotated pixels, which is tedious, time-consuming, expensive, and infeasible to acquire for many histology tasks. Thus, weakly-supervised semantic segmentation techniques are proposed to utilize weak supervision that is cheaper and quicker to acquire. In this paper, we propose SegGini, a weakly supervised segmentation method using graphs, that can utilize weak multiplex annotations, i.e. inexact and incomplete annotations, to segment arbitrary and large images, scaling from tissue microarray (TMA) to whole slide image (WSI). Formally, SegGini constructs a tissue-graph representation for an input histology image, where the graph nodes depict tissue regions. Then, it performs weakly-supervised segmentation via node classification by using inexact image-level labels, incomplete scribbles, or both. We evaluated SegGini on two public prostate cancer datasets containing TMAs and WSIs. Our method achieved state-of-the-art segmentation performance on both datasets for various annotation settings while being comparable to a pathologist baseline.

Antoine Spahr, Behzad Bozorgtabar, Jean-Philippe Thiran

Detecting anomalies in musculoskeletal radiographs is of paramount importance for large-scale screening in the radiology workflow. Supervised deep networks take for granted a large number of annotations by radiologists, which is often prohibitively very time-consuming to acquire. Moreover, supervised systems are tailored to closed set scenarios, e.g., trained models suffer from overfitting to previously seen rare anomalies at training. Instead, our approach's rationale is to use task agnostic pretext tasks to leverage unlabeled data based on a cross-sample similarity measure. Besides, we formulate a complex distribution of data from normal class within our framework to avoid a potential bias on the side of anomalies. Through extensive experiments, we show that our method outperforms baselines across unsupervised and self-supervised anomaly detection settings on a real-world medical dataset, the MURA dataset. We also provide rich ablation studies to analyze each training stage's effect and loss terms on the final performance.

Guillaume Jaume, Pushpak Pati, Behzad Bozorgtabar et al.

Explainability of deep learning methods is imperative to facilitate their clinical adoption in digital pathology. However, popular deep learning methods and explainability techniques (explainers) based on pixel-wise processing disregard biological entities' notion, thus complicating comprehension by pathologists. In this work, we address this by adopting biological entity-based graph processing and graph explainers enabling explanations accessible to pathologists. In this context, a major challenge becomes to discern meaningful explainers, particularly in a standardized and quantifiable fashion. To this end, we propose herein a set of novel quantitative metrics based on statistics of class separability using pathologically measurable concepts to characterize graph explainers. We employ the proposed metrics to evaluate three types of graph explainers, namely the layer-wise relevance propagation, gradient-based saliency, and graph pruning approaches, to explain Cell-Graph representations for Breast Cancer Subtyping. The proposed metrics are also applicable in other domains by using domain-specific intuitive concepts. We validate the qualitative and quantitative findings on the BRACS dataset, a large cohort of breast cancer RoIs, by expert pathologists.

Behzad Bozorgtabar, Dwarikanath Mahapatra, Guillaume Vray et al.

Deep anomaly detection models using a supervised mode of learning usually work under a closed set assumption and suffer from overfitting to previously seen rare anomalies at training, which hinders their applicability in a real scenario. In addition, obtaining annotations for X-rays is very time consuming and requires extensive training of radiologists. Hence, training anomaly detection in a fully unsupervised or self-supervised fashion would be advantageous, allowing a significant reduction of time spent on the report by radiologists. In this paper, we present SALAD, an end-to-end deep self-supervised methodology for anomaly detection on X-Ray images. The proposed method is based on an optimization strategy in which a deep neural network is encouraged to represent prototypical local patterns of the normal data in the embedding space. During training, we record the prototypical patterns of normal training samples via a memory bank. Our anomaly score is then derived by measuring similarity to a weighted combination of normal prototypical patterns within a memory bank without using any anomalous patterns. We present extensive experiments on the challenging NIH Chest X-rays and MURA dataset, which indicate that our algorithm improves state-of-the-art methods by a wide margin.

Dwarikanath Mahapatra, Behzad Bozorgtabar, Jean-Philippe Thiran et al.

Although generative adversarial network (GAN) based style transfer is state of the art in histopathology color-stain normalization, they do not explicitly integrate structural information of tissues. We propose a self-supervised approach to incorporate semantic guidance into a GAN based stain normalization framework and preserve detailed structural information. Our method does not require manual segmentation maps which is a significant advantage over existing methods. We integrate semantic information at different layers between a pre-trained semantic network and the stain color normalization network. The proposed scheme outperforms other color normalization methods leading to better classification and segmentation performance.

Christian Abbet, Inti Zlobec, Behzad Bozorgtabar et al.

With the long-term rapid increase in incidences of colorectal cancer (CRC), there is an urgent clinical need to improve risk stratification. The conventional pathology report is usually limited to only a few histopathological features. However, most of the tumor microenvironments used to describe patterns of aggressive tumor behavior are ignored. In this work, we aim to learn histopathological patterns within cancerous tissue regions that can be used to improve prognostic stratification for colorectal cancer. To do so, we propose a self-supervised learning method that jointly learns a representation of tissue regions as well as a metric of the clustering to obtain their underlying patterns. These histopathological patterns are then used to represent the interaction between complex tissues and predict clinical outcomes directly. We furthermore show that the proposed approach can benefit from linear predictors to avoid overfitting in patient outcomes predictions. To this end, we introduce a new well-characterized clinicopathological dataset, including a retrospective collective of 374 patients, with their survival time and treatment information. Histomorphological clusters obtained by our method are evaluated by training survival models. The experimental results demonstrate statistically significant patient stratification, and our approach outperformed the state-of-the-art deep clustering methods.

Mohammad Saeed Rad, Thomas Yu, Claudiu Musat et al.

Super-resolution (SR) has traditionally been based on pairs of high-resolution images (HR) and their low-resolution (LR) counterparts obtained artificially with bicubic downsampling. However, in real-world SR, there is a large variety of realistic image degradations and analytically modeling these realistic degradations can prove quite difficult. In this work, we propose to handle real-world SR by splitting this ill-posed problem into two comparatively more well-posed steps. First, we train a network to transform real LR images to the space of bicubically downsampled images in a supervised manner, by using both real LR/HR pairs and synthetic pairs. Second, we take a generic SR network trained on bicubically downsampled images to super-resolve the transformed LR image. The first step of the pipeline addresses the problem by registering the large variety of degraded images to a common, well understood space of images. The second step then leverages the already impressive performance of SR on bicubically downsampled images, sidestepping the issues of end-to-end training on datasets with many different image degradations. We demonstrate the effectiveness of our proposed method by comparing it to recent methods in real-world SR and show that our proposed approach outperforms the state-of-the-art works in terms of both qualitative and quantitative results, as well as results of an extensive user study conducted on several real image datasets.

Dwarikanath Mahapatra, Behzad Bozorgtabar, Jean-Philippe Thiran et al.

Medical image segmentation is an important task for computer aided diagnosis. Pixelwise manual annotations of large datasets require high expertise and is time consuming. Conventional data augmentations have limited benefit by not fully representing the underlying distribution of the training set, thus affecting model robustness when tested on images captured from different sources. Prior work leverages synthetic images for data augmentation ignoring the interleaved geometric relationship between different anatomical labels. We propose improvements over previous GAN-based medical image synthesis methods by jointly encoding the intrinsic relationship of geometry and shape. Latent space variable sampling results in diverse generated images from a base image and improves robustness. Given those augmented images generated by our method, we train the segmentation network to enhance the segmentation performance of retinal optical coherence tomography (OCT) images. The proposed method outperforms state-of-the-art segmentation methods on the public RETOUCH dataset having images captured from different acquisition procedures. Ablation studies and visual analysis also demonstrate benefits of integrating geometry and diversity.

Anca-Nicoleta Ciubotaru, Arnout Devos, Behzad Bozorgtabar et al.

Most of the existing deep neural nets on automatic facial expression recognition focus on a set of predefined emotion classes, where the amount of training data has the biggest impact on performance. However, in the standard setting over-parameterised neural networks are not amenable for learning from few samples as they can quickly over-fit. In addition, these approaches do not have such a strong generalisation ability to identify a new category, where the data of each category is too limited and significant variations exist in the expression within the same semantic category. We embrace these challenges and formulate the problem as a low-shot learning, where once the base classifier is deployed, it must rapidly adapt to recognise novel classes using a few samples. In this paper, we revisit and compare existing few-shot learning methods for the low-shot facial expression recognition in terms of their generalisation ability via episode-training. In particular, we extend our analysis on the cross-domain generalisation, where training and test tasks are not drawn from the same distribution. We demonstrate the efficacy of low-shot learning methods through extensive experiments.

Mohammad Saeed Rad, Behzad Bozorgtabar, Urs-Viktor Marti et al.

By benefiting from perceptual losses, recent studies have improved significantly the performance of the super-resolution task, where a high-resolution image is resolved from its low-resolution counterpart. Although such objective functions generate near-photorealistic results, their capability is limited, since they estimate the reconstruction error for an entire image in the same way, without considering any semantic information. In this paper, we propose a novel method to benefit from perceptual loss in a more objective way. We optimize a deep network-based decoder with a targeted objective function that penalizes images at different semantic levels using the corresponding terms. In particular, the proposed method leverages our proposed OBB (Object, Background and Boundary) labels, generated from segmentation labels, to estimate a suitable perceptual loss for boundaries, while considering texture similarity for backgrounds. We show that our proposed approach results in more realistic textures and sharper edges, and outperforms other state-of-the-art algorithms in terms of both qualitative results on standard benchmarks and results of extensive user studies.

Mohammad Saeed Rad, Behzad Bozorgtabar, Claudiu Musat et al.

Despite significant progress toward super resolving more realistic images by deeper convolutional neural networks (CNNs), reconstructing fine and natural textures still remains a challenging problem. Recent works on single image super resolution (SISR) are mostly based on optimizing pixel and content wise similarity between recovered and high-resolution (HR) images and do not benefit from recognizability of semantic classes. In this paper, we introduce a novel approach using categorical information to tackle the SISR problem; we present a decoder architecture able to extract and use semantic information to super-resolve a given image by using multitask learning, simultaneously for image super-resolution and semantic segmentation. To explore categorical information during training, the proposed decoder only employs one shared deep network for two task-specific output layers. At run-time only layers resulting HR image are used and no segmentation label is required. Extensive perceptual experiments and a user study on images randomly selected from COCO-Stuff dataset demonstrate the effectiveness of our proposed method and it outperforms the state-of-the-art methods.

Omid Abdollahi Aghdam, Behzad Bozorgtabar, Hazım Kemal Ekenel et al.

State-of-the-art deep face recognition approaches report near perfect performance on popular benchmarks, e.g., Labeled Faces in the Wild. However, their performance deteriorates significantly when they are applied on low quality images, such as those acquired by surveillance cameras. A further challenge for low resolution face recognition for surveillance applications is the matching of recorded low resolution probe face images with high resolution reference images, which could be the case in watchlist scenarios. In this paper, we have addressed these problems and investigated the factors that would contribute to the identification performance of the state-of-the-art deep face recognition models when they are applied to low resolution face recognition under mismatched conditions. We have observed that the following factors affect performance in a positive way: appearance variety and resolution distribution of the training dataset, resolution matching between the gallery and probe images, and the amount of information included in the probe images. By leveraging this information, we have utilized deep face models trained on MS-Celeb-1M and fine-tuned on VGGFace2 dataset and achieved state-of-the-art accuracies on the SCFace and ICB-RW benchmarks, even without using any training data from the datasets of these benchmarks.

Behzad Bozorgtabar, Mohammad Saeed Rad, Hazim Kemal Ekenel et al.

Cross-domain synthesizing realistic faces to learn deep models has attracted increasing attention for facial expression analysis as it helps to improve the performance of expression recognition accuracy despite having small number of real training images. However, learning from synthetic face images can be problematic due to the distribution discrepancy between low-quality synthetic images and real face images and may not achieve the desired performance when the learned model applies to real world scenarios. To this end, we propose a new attribute guided face image synthesis to perform a translation between multiple image domains using a single model. In addition, we adopt the proposed model to learn from synthetic faces by matching the feature distributions between different domains while preserving each domain's characteristics. We evaluate the effectiveness of the proposed approach on several face datasets on generating realistic face images. We demonstrate that the expression recognition performance can be enhanced by benefiting from our face synthesis model. Moreover, we also conduct experiments on a near-infrared dataset containing facial expression videos of drivers to assess the performance using in-the-wild data for driver emotion recognition.

Behzad Bozorgtabar, Mohammad Saeed Rad, Hazım Kemal Ekenel et al.

Attribute guided face image synthesis aims to manipulate attributes on a face image. Most existing methods for image-to-image translation can either perform a fixed translation between any two image domains using a single attribute or require training data with the attributes of interest for each subject. Therefore, these methods could only train one specific model for each pair of image domains, which limits their ability in dealing with more than two domains. Another disadvantage of these methods is that they often suffer from the common problem of mode collapse that degrades the quality of the generated images. To overcome these shortcomings, we propose attribute guided face image generation method using a single model, which is capable to synthesize multiple photo-realistic face images conditioned on the attributes of interest. In addition, we adopt the proposed model to increase the realism of the simulated face images while preserving the face characteristics. Compared to existing models, synthetic face images generated by our method present a good photorealistic quality on several face datasets. Finally, we demonstrate that generated facial images can be used for synthetic data augmentation, and improve the performance of the classifier used for facial expression recognition.

Behzad Bozorgtabar, Dwarikanath Mahapatra, Hendrik von Teng et al.

Training robust deep learning (DL) systems for disease detection from medical images is challenging due to limited images covering different disease types and severity. The problem is especially acute, where there is a severe class imbalance. We propose an active learning (AL) framework to select most informative samples for training our model using a Bayesian neural network. Informative samples are then used within a novel class aware generative adversarial network (CAGAN) to generate realistic chest xray images for data augmentation by transferring characteristics from one class label to another. Experiments show our proposed AL framework is able to achieve state-of-the-art performance by using about $35\%$ of the full dataset, thus saving significant time and effort over conventional methods.

Dwarikanath Mahapatra, Behzad Bozorgtabar

Anatomical landmark segmentation and pathology localization are important steps in automated analysis of medical images. They are particularly challenging when the anatomy or pathology is small, as in retinal images and cardiac MRI, or when the image is of low quality due to device acquisition parameters as in magnetic resonance (MR) scanners. We propose an image super-resolution method using progressive generative adversarial networks (P-GAN) that can take as input a low-resolution image and generate a high resolution image of desired scaling factor. The super resolved images can be used for more accurate detection of landmarks and pathology. Our primary contribution is in proposing a multistage model where the output image quality of one stage is progressively improved in the next stage by using a triplet loss function. The triplet loss enables stepwise image quality improvement by using the output of the previous stage as the baseline. This facilitates generation of super resolved images of high scaling factor while maintaining good image quality. Experimental results for image super-resolution show that our proposed multistage P-GAN outperforms competing methods and baseline GAN.

Guillaume Jaume, Behzad Bozorgtabar, Hazim Kemal Ekenel et al.

We introduce a new scene graph generation method called image-level attentional context modeling (ILAC). Our model includes an attentional graph network that effectively propagates contextual information across the graph using image-level features. Whereas previous works use an object-centric context, we build an image-level context agent to encode the scene properties. The proposed method comprises a single-stream network that iteratively refines the scene graph with a nested graph neural network. We demonstrate that our approach achieves competitive performance with the state-of-the-art for scene graph generation on the Visual Genome dataset, while requiring fewer parameters than other methods. We also show that ILAC can improve regular object detectors by incorporating relational image-level information.

Dwarikanath Mahapatra, Behzad Bozorgtabar, Jean-Philippe Thiran et al.

Training robust deep learning (DL) systems for medical image classification or segmentation is challenging due to limited images covering different disease types and severity. We propose an active learning (AL) framework to select most informative samples and add to the training data. We use conditional generative adversarial networks (cGANs) to generate realistic chest xray images with different disease characteristics by conditioning its generation on a real image sample. Informative samples to add to the training set are identified using a Bayesian neural network. Experiments show our proposed AL framework is able to achieve state of the art performance by using about 35% of the full dataset, thus saving significant time and effort over conventional methods.

Dwarikanath Mahapatra, Behzad Bozorgtabar

We propose an image super resolution(ISR) method using generative adversarial networks (GANs) that takes a low resolution input fundus image and generates a high resolution super resolved (SR) image upto scaling factor of $16$. This facilitates more accurate automated image analysis, especially for small or blurred landmarks and pathologies. Local saliency maps, which define each pixel's importance, are used to define a novel saliency loss in the GAN cost function. Experimental results show the resulting SR images have perceptual quality very close to the original images and perform better than competing methods that do not weigh pixels according to their importance. When used for retinal vasculature segmentation, our SR images result in accuracy levels close to those obtained when using the original images.