Orcun Goksel

Kevin Thandiackal, Boqi Chen, Pushpak Pati et al.

Multiple Instance Learning (MIL) methods have become increasingly popular for classifying giga-pixel sized Whole-Slide Images (WSIs) in digital pathology. Most MIL methods operate at a single WSI magnification, by processing all the tissue patches. Such a formulation induces high computational requirements, and constrains the contextualization of the WSI-level representation to a single scale. A few MIL methods extend to multiple scales, but are computationally more demanding. In this paper, inspired by the pathological diagnostic process, we propose ZoomMIL, a method that learns to perform multi-level zooming in an end-to-end manner. ZoomMIL builds WSI representations by aggregating tissue-context information from multiple magnifications. The proposed method outperforms the state-of-the-art MIL methods in WSI classification on two large datasets, while significantly reducing the computational demands with regard to Floating-Point Operations (FLOPs) and processing time by up to 40x.

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.

Boqi Chen, Kevin Thandiackal, Pushpak Pati et al.

Deep learning models can achieve high accuracy when trained on large amounts of labeled data. However, real-world scenarios often involve several challenges: Training data may become available in installments, may originate from multiple different domains, and may not contain labels for training. Certain settings, for instance medical applications, often involve further restrictions that prohibit retention of previously seen data due to privacy regulations. In this work, to address such challenges, we study unsupervised segmentation in continual learning scenarios that involve domain shift. To that end, we introduce GarDA (Generative Appearance Replay for continual Domain Adaptation), a generative-replay based approach that can adapt a segmentation model sequentially to new domains with unlabeled data. In contrast to single-step unsupervised domain adaptation (UDA), continual adaptation to a sequence of domains enables leveraging and consolidation of information from multiple domains. Unlike previous approaches in incremental UDA, our method does not require access to previously seen data, making it applicable in many practical scenarios. We evaluate GarDA on two datasets with different organs and modalities, where it substantially outperforms existing techniques.

Alvaro Gomariz, Huanxiang Lu, Yun Yvonna Li et al.

Accurate segmentation of retinal fluids in 3D Optical Coherence Tomography images is key for diagnosis and personalized treatment of eye diseases. While deep learning has been successful at this task, trained supervised models often fail for images that do not resemble labeled examples, e.g. for images acquired using different devices. We hereby propose a novel semi-supervised learning framework for segmentation of volumetric images from new unlabeled domains. We jointly use supervised and contrastive learning, also introducing a contrastive pairing scheme that leverages similarity between nearby slices in 3D. In addition, we propose channel-wise aggregation as an alternative to conventional spatial-pooling aggregation for contrastive feature map projection. We evaluate our methods for domain adaptation from a (labeled) source domain to an (unlabeled) target domain, each containing images acquired with different acquisition devices. In the target domain, our method achieves a Dice coefficient 13.8% higher than SimCLR (a state-of-the-art contrastive framework), and leads to results comparable to an upper bound with supervised training in that domain. In the source domain, our model also improves the results by 5.4% Dice, by successfully leveraging information from many unlabeled images.

Kevin Thandiackal, Luigi Piccinelli, Pushpak Pati et al.

Methods for unsupervised domain adaptation (UDA) help to improve the performance of deep neural networks on unseen domains without any labeled data. Especially in medical disciplines such as histopathology, this is crucial since large datasets with detailed annotations are scarce. While the majority of existing UDA methods focus on the adaptation from a labeled source to a single unlabeled target domain, many real-world applications with a long life cycle involve more than one target domain. Thus, the ability to sequentially adapt to multiple target domains becomes essential. In settings where the data from previously seen domains cannot be stored, e.g., due to data protection regulations, the above becomes a challenging continual learning problem. To this end, we propose to use generative feature-driven image replay in conjunction with a dual-purpose discriminator that not only enables the generation of images with realistic features for replay, but also promotes feature alignment during domain adaptation. We evaluate our approach extensively on a sequence of three histopathological datasets for tissue-type classification, achieving state-of-the-art results. We present detailed ablation experiments studying our proposed method components and demonstrate a possible use-case of our continual UDA method for an unsupervised patch-based segmentation task given high-resolution tissue images.

Lin Zhang, Tiziano Portenier, Orcun Goksel

Photorealistic image generation from simulated label maps are necessitated in several contexts, such as for medical training in virtual reality. With conventional deep learning methods, this task requires images that are paired with semantic annotations, which typically are unavailable. We introduce a contrastive learning framework for generating photorealistic images from simulated label maps, by learning from unpaired sets of both. Due to potentially large scene differences between real images and label maps, existing unpaired image translation methods lead to artifacts of scene modification in synthesized images. We utilize simulated images as surrogate targets for a contrastive loss, while ensuring consistency by utilizing features from a reverse translation network. Our method enables bidirectional label-image translations, which is demonstrated in a variety of scenarios and datasets, including laparoscopy, ultrasound, and driving scenes. By comparing with state-of-the-art unpaired translation methods, our proposed method is shown to generate realistic and scene-accurate translations.

Swarnadip Chatterjee, Vladimir Basic, Arrigo Capitanio et al.

In computational cytology, detecting malignancy on whole-slide images is difficult because malignant cells are morphologically diverse yet vanishingly rare amid a vast background of normal cells. Accurate detection of these extremely rare malignant cells remains challenging due to large class imbalance and limited annotations. Conventional weakly supervised approaches, such as multiple instance learning (MIL), often fail to generalize at the instance level, especially when the fraction of malignant cells (witness rate) is exceedingly low. In this study, we explore the use of one-class representation learning techniques for detecting malignant cells in low-witness-rate scenarios. These methods are trained exclusively on slide-negative patches, without requiring any instance-level supervision. Specifically, we evaluate two OCC approaches, DSVDD and DROC, and compare them with FS-SIL, WS-SIL, and the recent ItS2CLR method. The one-class methods learn compact representations of normality and detect deviations at test time. Experiments on a publicly available bone marrow cytomorphology dataset (TCIA) and an in-house oral cancer cytology dataset show that DSVDD achieves state-of-the-art performance in instance-level abnormality ranking, particularly in ultra-low witness-rate regimes ($\leq 1\%$) and, in some cases, even outperforming fully supervised learning, which is typically not a practical option in whole-slide cytology due to the infeasibility of exhaustive instance-level annotations. DROC is also competitive under extreme rarity, benefiting from distribution-augmented contrastive learning. These findings highlight one-class representation learning as a robust and interpretable superior choice to MIL for malignant cell detection under extreme rarity.

Ture Hassler, Ida Åkerholm, Marcus Nordström et al.

Unet and its variations have been standard in semantic image segmentation, especially for computer assisted radiology. Current Unet architectures iteratively downsample spatial resolution while increasing channel dimensions to preserve information content. Such a structure demands a large memory footprint, limiting training batch sizes and increasing inference latency. Channel pruning compresses Unet architecture without accuracy loss, but requires lengthy optimization and may not generalize across tasks and datasets. By investigating Unet pruning, we hypothesize that the final structure is the crucial factor, not the channel selection strategy of pruning. Based on our observations, we propose a lean Unet architecture (LUnet) with a compact, flat hierarchy where channels are not doubled as resolution is halved. We evaluate on a public MRI dataset allowing comparable reporting, as well as on two internal CT datasets. We show that a state-of-the-art pruning solution (STAMP) mainly prunes from the layers with the highest number of channels. Comparatively, simply eliminating a random channel at the pruning-identified layer or at the largest layer achieves similar or better performance. Our proposed LUnet with fixed architectures and over 30 times fewer parameters achieves performance comparable to both conventional Unet counterparts and data-adaptively pruned networks. The proposed lean Unet with constant channel count across layers requires far fewer parameters while achieving performance superior to standard Unet for the same total number of parameters. Skip connections allow Unet bottleneck channels to be largely reduced, unlike standard encoder-decoder architectures requiring increased bottleneck channels for information propagation.

Roman Denkin, Orcun Goksel

Accurate speed-of-sound (SoS) estimation is crucial for ultrasound image formation, yet conventional systems often rely on an assumed value for imaging. While several methods exist for SoS estimation, they typically depend on complex physical models of acoustic propagation. We propose to leverage conventional image analysis techniques and metrics, as a novel and simple approach to estimate tissue SoS. We study eleven metrics in three categories for assessing image quality, image similarity and multi-frame variation, by testing them in numerical simulations and phantom experiments. Among single-frame image quality metrics, conventional Focus and our proposed Smoothed Threshold Tenengrad metrics achieved satisfactory accuracy, however only when applied to compounded images. Image quality metrics were largely surpassed by various image comparison metrics, which exhibited errors consistently under 8 m/s even applied to a single pair of images. Particularly, Mean Square Error is a computationally efficient alternative for global estimation. Mutual Information and Correlation are found to be robust to processing small image segments, making them suitable, e.g., for multi-layer SoS estimation. The above metrics do not require access to raw channel data as they can operate on post-beamformed data, and in the case of image quality metrics they can operate on B-mode images, given that the beamforming SoS can be controlled for beamforming using a multitude of values. These image analysis based SoS estimation methods offer a computationally efficient and data-accessible alternative to conventional physics-based methods, with potential extensions to layered or local SoS imaging.

Lingkai Zhu, Can Deniz Bezek, Orcun Goksel

In recent years, the increasing size of deep learning models and their growing demand for computational resources have drawn significant attention to the practice of pruning neural networks, while aiming to preserve their accuracy. In unstructured gradual pruning, which sparsifies a network by gradually removing individual network parameters until a targeted network sparsity is reached, recent works show that both gradient and weight magnitudes should be considered. In this work, we show that such mechanism, e.g., the order of prioritization and selection criteria, is essential. We introduce a gradient-first magnitude-next strategy for choosing the parameters to prune, and show that a fixed-rate subselection criterion between these steps works better, in contrast to the annealing approach in the literature. We validate this on CIFAR-10 dataset, with multiple randomized initializations on both VGG-19 and ResNet-50 network backbones, for pruning targets of 90, 95, and 98% sparsity and for both initially dense and 50% sparse networks. Our proposed fixed-rate gradient-first gradual pruning (FGGP) approach outperforms its state-of-the-art alternatives in most of the above experimental settings, even occasionally surpassing the upperbound of corresponding dense network results, and having the highest ranking across the considered experimental settings.

Alvaro Gomariz, Yusuke Kikuchi, Yun Yvonna Li et al.

Despite their effectiveness, current deep learning models face challenges with images coming from different domains with varying appearance and content. We introduce SegCLR, a versatile framework designed to segment images across different domains, employing supervised and contrastive learning simultaneously to effectively learn from both labeled and unlabeled data. We demonstrate the superior performance of SegCLR through a comprehensive evaluation involving three diverse clinical datasets of 3D retinal Optical Coherence Tomography (OCT) images, for the slice-wise segmentation of fluids with various network configurations and verification across 10 different network initializations. In an unsupervised domain adaptation context, SegCLR achieves results on par with a supervised upper-bound model trained on the intended target domain. Notably, we discover that the segmentation performance of SegCLR framework is marginally impacted by the abundance of unlabeled data from the target domain, thereby we also propose an effective domain generalization extension of SegCLR, known also as zero-shot domain adaptation, which eliminates the need for any target domain information. This shows that our proposed addition of contrastive loss in standard supervised training for segmentation leads to superior models, inherently more generalizable to both in- and out-of-domain test data. We additionally propose a pragmatic solution for SegCLR deployment in realistic scenarios with multiple domains containing labeled data. Accordingly, our framework pushes the boundaries of deep-learning based segmentation in multi-domain applications, regardless of data availability - labeled, unlabeled, or nonexistent.

Sonia Laguna, Lin Zhang, Can Deniz Bezek et al.

Speed-of-sound (SoS) is a biomechanical characteristic of tissue, and its imaging can provide a promising biomarker for diagnosis. Reconstructing SoS images from ultrasound acquisitions can be cast as a limited-angle computed-tomography problem, with Variational Networks being a promising model-based deep learning solution. Some acquired data frames may, however, get corrupted by noise due to, e.g., motion, lack of contact, and acoustic shadows, which in turn negatively affects the resulting SoS reconstructions. We propose to use the uncertainty in SoS reconstructions to attribute trust to each individual acquired frame. Given multiple acquisitions, we then use an uncertainty based automatic selection among these retrospectively, to improve diagnostic decisions. We investigate uncertainty estimation based on Monte Carlo Dropout and Bayesian Variational Inference. We assess our automatic frame selection method for differential diagnosis of breast cancer, distinguishing between benign fibroadenoma and malignant carcinoma. We evaluate 21 lesions classified as BI-RADS~4, which represents suspicious cases for probable malignancy. The most trustworthy frame among four acquisitions of each lesion was identified using uncertainty based criteria. Selecting a frame informed by uncertainty achieved an area under curve of 76% and 80% for Monte Carlo Dropout and Bayesian Variational Inference, respectively, superior to any uncertainty-uninformed baselines with the best one achieving 64%. A novel use of uncertainty estimation is proposed for selecting one of multiple data acquisitions for further processing and decision making.

Can Deniz Bezek, Orcun Goksel

Medical imaging aims to recover underlying tissue properties, using inexact (simplified/linearized) imaging models and often from inaccurate and incomplete measurements. Analytical reconstruction methods rely on hand-crafted regularization, sensitive to noise assumptions and parameter tuning. Among deep learning alternatives, plug-and-play (PnP) approaches learn regularization while incorporating imaging physics during inference, outperforming purely data-driven methods. The performance of all these approaches, however, still strongly depends on measurement quality and imaging model accuracy. In this work, we propose DenOiS, a framework that denoises both input observations and resulting solution in their respective domains. It consists of an observation refinement strategy that corrects degraded measurements while compensating for imaging model simplifications, and a diffusion-based PnP reconstruction approach that remains robust under missing measurements. DenOiS enables generalization to real data from training only in simulations, resulting in high-fidelity image reconstruction with noisy observations and inexact imaging models. We demonstrate this for speed-of-sound imaging as a challenging setting of quantitative ultrasound image reconstruction.

Kevin Thandiackal, Tiziano Portenier, Andrea Giovannini et al.

Neural networks are prone to catastrophic forgetting when trained incrementally on different tasks. Popular incremental learning methods mitigate such forgetting by retaining a subset of previously seen samples and replaying them during the training on subsequent tasks. However, this is not always possible, e.g., due to data protection regulations. In such restricted scenarios, one can employ generative models to replay either artificial images or hidden features to a classifier. In this work, we propose Genifer (GENeratIve FEature-driven image Replay), where a generative model is trained to replay images that must induce the same hidden features as real samples when they are passed through the classifier. Our technique therefore incorporates the benefits of both image and feature replay, i.e.: (1) unlike conventional image replay, our generative model explicitly learns the distribution of features that are relevant for classification; (2) in contrast to feature replay, our entire classifier remains trainable; and (3) we can leverage image-space augmentations, which increase distillation performance while also mitigating overfitting during the training of the generative model. We show that Genifer substantially outperforms the previous state of the art for various settings on the CIFAR-100 and CUB-200 datasets.

Devavrat Tomar, Lin Zhang, Tiziano Portenier et al.

Interactive simulation of ultrasound imaging greatly facilitates sonography training. Although ray-tracing based methods have shown promising results, obtaining realistic images requires substantial modeling effort and manual parameter tuning. In addition, current techniques still result in a significant appearance gap between simulated images and real clinical scans. Herein we introduce a novel content-preserving image translation framework (ConPres) to bridge this appearance gap, while maintaining the simulated anatomical layout. We achieve this goal by leveraging both simulated images with semantic segmentations and unpaired in-vivo ultrasound scans. Our framework is based on recent contrastive unpaired translation techniques and we propose a regularization approach by learning an auxiliary segmentation-to-real image translation task, which encourages the disentanglement of content and style. In addition, we extend the generator to be class-conditional, which enables the incorporation of additional losses, in particular a cyclic consistency loss, to further improve the translation quality. Qualitative and quantitative comparisons against state-of-the-art unpaired translation methods demonstrate the superiority of our proposed framework.

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.

Alvaro Gomariz, Tiziano Portenier, César Nombela-Arrieta et al.

3D microscopy is key in the investigation of diverse biological systems, and the ever increasing availability of large datasets demands automatic cell identification methods that not only are accurate, but also can imply the uncertainty in their predictions to inform about potential errors and hence confidence in conclusions using them. While conventional deep learning methods often yield deterministic results, advances in deep Bayesian learning allow for accurate predictions with a probabilistic interpretation in numerous image classification and segmentation tasks. It is however nontrivial to extend such Bayesian methods to cell detection, which requires specialized learning frameworks. In particular, regression of density maps is a popular successful approach for extracting cell coordinates from local peaks in a postprocessing step, which hinders any meaningful probabilistic output. We herein propose a deep learning-based cell detection framework that can operate on large microscopy images and outputs desired probabilistic predictions by (i) integrating Bayesian techniques for the regression of uncertainty-aware density maps, where peak detection can be applied to generate cell proposals, and (ii) learning a mapping from the numerous proposals to a probabilistic space that is calibrated, i.e. accurately represents the chances of a successful prediction. Utilizing such calibrated predictions, we propose a probabilistic spatial analysis with Monte-Carlo sampling. We demonstrate this in revising an existing description of the distribution of a mesenchymal stromal cell type within the bone marrow, where our proposed methods allow us to reveal spatial patterns that are otherwise undetectable. Introducing such probabilistic analysis in quantitative microscopy pipelines will allow for reporting confidence intervals for testing biological hypotheses of spatial distributions.

Pushpak Pati, Guillaume Jaume, Antonio Foncubierta et al.

Cancer diagnosis, prognosis, and therapy response predictions from tissue specimens highly depend on the phenotype and topological distribution of constituting histological entities. Thus, adequate tissue representations for encoding histological entities is imperative for computer aided cancer patient care. To this end, several approaches have leveraged cell-graphs that encode cell morphology and organization to denote the tissue information. These allow for utilizing machine learning to map tissue representations to tissue functionality to help quantify their relationship. Though cellular information is crucial, it is incomplete alone to comprehensively characterize complex tissue structure. We herein treat the tissue as a hierarchical composition of multiple types of histological entities from fine to coarse level, capturing multivariate tissue information at multiple levels. We propose a novel multi-level hierarchical entity-graph representation of tissue specimens to model hierarchical compositions that encode histological entities as well as their intra- and inter-entity level interactions. Subsequently, a graph neural network is proposed to operate on the hierarchical entity-graph representation to map the tissue structure to tissue functionality. Specifically, for input histology images we utilize well-defined cells and tissue regions to build HierArchical Cell-to-Tissue (HACT) graph representations, and devise HACT-Net, a graph neural network, to classify such HACT representations. As part of this work, we introduce the BReAst Carcinoma Subtyping (BRACS) dataset, a large cohort of H&E stained breast tumor images, to evaluate our proposed methodology against pathologists and state-of-the-art approaches. Through comparative assessment and ablation studies, our method is demonstrated to yield superior classification results compared to alternative methods as well as pathologists.

Alvaro Gomariz, Raphael Egli, Tiziano Portenier et al.

Fluorescence microscopy images contain several channels, each indicating a marker staining the sample. Since many different marker combinations are utilized in practice, it has been challenging to apply deep learning based segmentation models, which expect a predefined channel combination for all training samples as well as at inference for future application. Recent work circumvents this problem using a modality attention approach to be effective across any possible marker combination. However, for combinations that do not exist in a labeled training dataset, one cannot have any estimation of potential segmentation quality if that combination is encountered during inference. Without this, not only one lacks quality assurance but one also does not know where to put any additional imaging and labeling effort. We herein propose a method to estimate segmentation quality on unlabeled images by (i) estimating both aleatoric and epistemic uncertainties of convolutional neural networks for image segmentation, and (ii) training a Random Forest model for the interpretation of uncertainty features via regression to their corresponding segmentation metrics. Additionally, we demonstrate that including these uncertainty measures during training can provide an improvement on segmentation performance.

Lin Zhang, Tiziano Portenier, Orcun Goksel

Purpose. Given the high level of expertise required for navigation and interpretation of ultrasound images, computational simulations can facilitate the training of such skills in virtual reality. With ray-tracing based simulations, realistic ultrasound images can be generated. However, due to computational constraints for interactivity, image quality typically needs to be compromised. Methods. We propose herein to bypass any rendering and simulation process at interactive time, by conducting such simulations during a non-time-critical offline stage and then learning image translation from cross-sectional model slices to such simulated frames. We use a generative adversarial framework with a dedicated generator architecture and input feeding scheme, which both substantially improve image quality without increase in network parameters. Integral attenuation maps derived from cross-sectional model slices, texture-friendly strided convolutions, providing stochastic noise and input maps to intermediate layers in order to preserve locality are all shown herein to greatly facilitate such translation task. Results. Given several quality metrics, the proposed method with only tissue maps as input is shown to provide comparable or superior results to a state-of-the-art that uses additional images of low-quality ultrasound renderings. An extensive ablation study shows the need and benefits from the individual contributions utilized in this work, based on qualitative examples and quantitative ultrasound similarity metrics. To that end, a local histogram statistics based error metric is proposed and demonstrated for visualization of local dissimilarities between ultrasound images.

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.

Alvaro Gomariz, Tiziano Portenier, Patrick M. Helbling et al.

Fluorescence microscopy allows for a detailed inspection of cells, cellular networks, and anatomical landmarks by staining with a variety of carefully-selected markers visualized as color channels. Quantitative characterization of structures in acquired images often relies on automatic image analysis methods. Despite the success of deep learning methods in other vision applications, their potential for fluorescence image analysis remains underexploited. One reason lies in the considerable workload required to train accurate models, which are normally specific for a given combination of markers, and therefore applicable to a very restricted number of experimental settings. We herein propose Marker Sampling and Excite, a neural network approach with a modality sampling strategy and a novel attention module that together enable (i) flexible training with heterogeneous datasets with combinations of markers and (ii) successful utility of learned models on arbitrary subsets of markers prospectively. We show that our single neural network solution performs comparably to an upper bound scenario where an ensemble of many networks is naïvely trained for each possible marker combination separately. In addition, we demonstrate the feasibility of this framework in high-throughput biological analysis by revising a recent quantitative characterization of bone marrow vasculature in 3D confocal microscopy datasets and further confirm the validity of our approach on an additional, significantly different dataset of microvessels in fetal liver tissues. Not only can our work substantially ameliorate the use of deep learning in fluorescence microscopy analysis, but it can also be utilized in other fields with incomplete data acquisitions and missing modalities.

Emanuel Joos, Fabien Péan, Orcun Goksel

To diagnose, plan, and treat musculoskeletal pathologies, understanding and reproducing muscle recruitment for complex movements is essential. With muscle activations for movements often being highly redundant, nonlinear, and time dependent, machine learning can provide a solution for their modeling and control for anatomy-specific musculoskeletal simulations. Sophisticated biomechanical simulations often require specialized computational environments, being numerically complex and slow, hindering their integration with typical deep learning frameworks. In this work, a deep reinforcement learning (DRL) based inverse dynamics controller is trained to control muscle activations of a biomechanical model of the human shoulder. In a generalizable end-to-end fashion, muscle activations are learned given current and desired position-velocity pairs. A customized reward functions for trajectory control is introduced, enabling straightforward extension to additional muscles and higher degrees of freedom. Using the biomechanical model, multiple episodes are simulated on a cluster simultaneously using the evolving neural models of the DRL being trained. Results are presented for a single-axis motion control of shoulder abduction for the task of following randomly generated angular trajectories.

Pushpak Pati, Guillaume Jaume, Lauren Alisha Fernandes et al.

Cancer diagnosis, prognosis, and therapeutic response prediction are heavily influenced by the relationship between the histopathological structures and the function of the tissue. Recent approaches acknowledging the structure-function relationship, have linked the structural and spatial patterns of cell organization in tissue via cell-graphs to tumor grades. Though cell organization is imperative, it is insufficient to entirely represent the histopathological structure. We propose a novel hierarchical cell-to-tissue-graph (HACT) representation to improve the structural depiction of the tissue. It consists of a low-level cell-graph, capturing cell morphology and interactions, a high-level tissue-graph, capturing morphology and spatial distribution of tissue parts, and cells-to-tissue hierarchies, encoding the relative spatial distribution of the cells with respect to the tissue distribution. Further, a hierarchical graph neural network (HACT-Net) is proposed to efficiently map the HACT representations to histopathological breast cancer subtypes. We assess the methodology on a large set of annotated tissue regions of interest from H\&E stained breast carcinoma whole-slides. Upon evaluation, the proposed method outperformed recent convolutional neural network and graph neural network approaches for breast cancer multi-class subtyping. The proposed entity-based topological analysis is more inline with the pathological diagnostic procedure of the tissue. It provides more command over the tissue modelling, therefore encourages the further inclusion of pathological priors into task-specific tissue representation.

Guillaume Jaume, Pushpak Pati, Antonio Foncubierta-Rodriguez et al.

Explainability of machine learning (ML) techniques in digital pathology (DP) is of great significance to facilitate their wide adoption in clinics. Recently, graph techniques encoding relevant biological entities have been employed to represent and assess DP images. Such paradigm shift from pixel-wise to entity-wise analysis provides more control over concept representation. In this paper, we introduce a post-hoc explainer to derive compact per-instance explanations emphasizing diagnostically important entities in the graph. Although we focus our analyses to cells and cellular interactions in breast cancer subtyping, the proposed explainer is generic enough to be extended to other topological representations in DP. Qualitative and quantitative analyses demonstrate the efficacy of the explainer in generating comprehensive and compact explanations.

Tiziano Portenier, Siavash Bigdeli, Orcun Goksel

We present a novel texture synthesis framework, enabling the generation of infinite, high-quality 3D textures given a 2D exemplar image. Inspired by recent advances in natural texture synthesis, we train deep neural models to generate textures by non-linearly combining learned noise frequencies. To achieve a highly realistic output conditioned on an exemplar patch, we propose a novel loss function that combines ideas from both style transfer and generative adversarial networks. In particular, we train the synthesis network to match the Gram matrices of deep features from a discriminator network. In addition, we propose two architectural concepts and an extrapolation strategy that significantly improve generalization performance. In particular, we inject both model input and condition into hidden network layers by learning to scale and bias hidden activations. Quantitative and qualitative evaluations on a diverse set of exemplars motivate our design decisions and show that our system performs superior to previous state of the art. Finally, we conduct a user study that confirms the benefits of our framework.

Melanie Bernhardt, Valery Vishnevskiy, Richard Rau et al.

Speed-of-sound has been shown as a potential biomarker for breast cancer imaging, successfully differentiating malignant tumors from benign ones. Speed-of-sound images can be reconstructed from time-of-flight measurements from ultrasound images acquired using conventional handheld ultrasound transducers. Variational Networks (VN) have recently been shown to be a potential learning-based approach for optimizing inverse problems in image reconstruction. Despite earlier promising results, these methods however do not generalize well from simulated to acquired data, due to the domain shift. In this work, we present for the first time a VN solution for a pulse-echo SoS image reconstruction problem using diverging waves with conventional transducers and single-sided tissue access. This is made possible by incorporating simulations with varying complexity into training. We use loop unrolling of gradient descent with momentum, with an exponentially weighted loss of outputs at each unrolled iteration in order to regularize training. We learn norms as activation functions regularized to have smooth forms for robustness to input distribution variations. We evaluate reconstruction quality on ray-based and full-wave simulations as well as on tissue-mimicking phantom data, in comparison to a classical iterative (L-BFGS) optimization of this image reconstruction problem. We show that the proposed regularization techniques combined with multi-source domain training yield substantial improvements in the domain adaptation capabilities of VN, reducing median RMSE by 54% on a wave-based simulation dataset compared to the baseline VN. We also show that on data acquired from a tissue-mimicking breast phantom the proposed VN provides improved reconstruction in 12 milliseconds.

Lin Zhang, Tiziano Portenier, Christoph Paulus et al.

Ultrasound simulation based on ray tracing enables the synthesis of highly realistic images. It can provide an interactive environment for training sonographers as an educational tool. However, due to high computational demand, there is a trade-off between image quality and interactivity, potentially leading to sub-optimal results at interactive rates. In this work we introduce a deep learning approach based on adversarial training that mitigates this trade-off by improving the quality of simulated images with constant computation time. An image-to-image translation framework is utilized to translate low quality images into high quality versions. To incorporate anatomical information potentially lost in low quality images, we additionally provide segmentation maps to image translation. Furthermore, we propose to leverage information from acoustic attenuation maps to better preserve acoustic shadows and directional artifacts, an invaluable feature for ultrasound image interpretation. The proposed method yields an improvement of 7.2% in Fréchet Inception Distance and 8.9% in patch-based Kullback-Leibler divergence.

Lin Zhang, Valery Vishnevskiy, Orcun Goksel

Simulation-based ultrasound training can be an essential educational tool. Realistic ultrasound image appearance with typical speckle texture can be modeled as convolution of a point spread function with point scatterers representing tissue microstructure. Such scatterer distribution, however, is in general not known and its estimation for a given tissue type is fundamentally an ill-posed inverse problem. In this paper, we demonstrate a convolutional neural network approach for probabilistic scatterer estimation from observed ultrasound data. We herein propose to impose a known statistical distribution on scatterers and learn the mapping between ultrasound image and distribution parameter map by training a convolutional neural network on synthetic images. In comparison with several existing approaches, we demonstrate in numerical simulations and with in-vivo images that the synthesized images from scatterer representations estimated with our approach closely match the observations with varying acquisition parameters such as compression and rotation of the imaged domain.

Pushpak Pati, Antonio Foncubierta-Rodriguez, Orcun Goksel et al.

Mitotic counting is a vital prognostic marker of tumor proliferation in breast cancer. Deep learning-based mitotic detection is on par with pathologists, but it requires large labeled data for training. We propose a deep classification framework for enhancing mitosis detection by leveraging class label information, via softmax loss, and spatial distribution information among samples, via distance metric learning. We also investigate strategies towards steadily providing informative samples to boost the learning. The efficacy of the proposed framework is established through evaluation on ICPR 2012 and AMIDA 2013 mitotic data. Our framework significantly improves the detection with small training data and achieves on par or superior performance compared to state-of-the-art methods for using the entire training data.

Firat Ozdemir, Christine Tanner, Orcun Goksel

Bone surface delineation in ultrasound is of interest due to its potential in diagnosis, surgical planning, and post-operative follow-up in orthopedics, as well as the potential of using bones as anatomical landmarks in surgical navigation. We herein propose a method to encode the physics of ultrasound propagation into a factor graph formulation for the purpose of bone surface delineation. In this graph structure, unary node potentials encode the local likelihood for being a soft tissue or acoustic-shadow (behind bone surface) region, both learned through image descriptors. Pair-wise edge potentials encode ultrasound propagation constraints of bone surfaces given their large acoustic-impedance difference. We evaluate the proposed method in comparison with four earlier approaches, on in-vivo ultrasound images collected from dorsal and volar views of the forearm. The proposed method achieves an average root-mean-square error and symmetric Hausdorff distance of 0.28mm and 1.78mm, respectively. It detects 99.9% of the annotated bone surfaces with a mean scanline error (distance to annotations) of 0.39mm.

Firat Ozdemir, Zixuan Peng, Philipp Fuernstahl et al.

Segmentation of anatomical structures is a fundamental image analysis task for many applications in the medical field. Deep learning methods have been shown to perform well, but for this purpose large numbers of manual annotations are needed in the first place, which necessitate prohibitive levels of resources that are often unavailable. In an active learning framework of selecting informed samples for manual labeling, expert clinician time for manual annotation can be optimally utilized, enabling the establishment of large labeled datasets for machine learning. In this paper, we propose a novel method that combines representativeness with uncertainty in order to estimate ideal samples to be annotated, iteratively from a given dataset. Our novel representativeness metric is based on Bayesian sampling, by using information-maximizing autoencoders. We conduct experiments on a shoulder magnetic resonance imaging (MRI) dataset for the segmentation of four musculoskeletal tissue classes. Quantitative results show that the annotation of representative samples selected by our proposed querying method yields an improved segmentation performance at each active learning iteration, compared to a baseline method that also employs uncertainty and representativeness metrics. For instance, with only 10% of the dataset annotated, our method reaches within 5% of Dice score expected from the upper bound scenario of all the dataset given as annotated (an impractical scenario due to resource constraints), and this gap drops down to a mere 2% when less than a fifth of the dataset samples are annotated. Such active learning approach to selecting samples to annotate enables an optimal use of the expert clinician time, being often the bottleneck in realizing machine learning solutions in medicine.

Valery Vishnevskiy, Richard Rau, Orcun Goksel

Tomographic image reconstruction is relevant for many medical imaging modalities including X-ray, ultrasound (US) computed tomography (CT) and photoacoustics, for which the access to full angular range tomographic projections might be not available in clinical practice due to physical or time constraints. Reconstruction from incomplete data in low signal-to-noise ratio regime is a challenging and ill-posed inverse problem that usually leads to unsatisfactory image quality. While informative image priors may be learned using generic deep neural network architectures, the artefacts caused by an ill-conditioned design matrix often have global spatial support and cannot be efficiently filtered out by means of convolutions. In this paper we propose to learn an inverse mapping in an end-to-end fashion via unrolling optimization iterations of a prototypical reconstruction algorithm. We herein introduce a network architecture that performs filtering jointly in both sinogram and spatial domains. To efficiently train such deep network we propose a novel regularization approach based on deep exponential weighting. Experiments on US and X-ray CT data show that our proposed method is qualitatively and quantitatively superior to conventional non-linear reconstruction methods as well as state-of-the-art deep networks for image reconstruction. Fast inference time of the proposed algorithm allows for sophisticated reconstructions in real-time critical settings, demonstrated with US SoS imaging of an ex vivo bovine phantom.

Yuanhao Gong, Orcun Goksel

In image processing tasks, spatial priors are essential for robust computations, regularization, algorithmic design and Bayesian inference. In this paper, we introduce weighted mean curvature (WMC) as a novel image prior and present an efficient computation scheme for its discretization in practical image processing applications. We first demonstrate the favorable properties of WMC, such as sampling invariance, scale invariance, and contrast invariance with Gaussian noise model; and we show the relation of WMC to area regularization. We further propose an efficient computation scheme for discretized WMC, which is demonstrated herein to process over 33.2 giga-pixels/second on GPU. This scheme yields itself to a convolutional neural network representation. Finally, WMC is evaluated on synthetic and real images, showing its superiority quantitatively to total-variation and mean curvature.

Andrawes Al Bahou, Christine Tanner, Orcun Goksel

Computational simulation of ultrasound (US) echography is essential for training sonographers. Realistic simulation of US interaction with microscopic tissue structures is often modeled by a tissue representation in the form of point scatterers, convolved with a spatially varying point spread function. This yields a realistic US B-mode speckle texture, given that a scatterer representation for a particular tissue type is readily available. This is often not the case and scatterers are nontrivial to determine. In this work we propose to estimate scatterer maps from sample US B-mode images of a tissue, by formulating this inverse mapping problem as image translation, where we learn the mapping with Generative Adversarial Networks, using a US simulation software for training. We demonstrate robust reconstruction results, invariant to US viewing and imaging settings such as imaging direction and center frequency. Our method is shown to generalize beyond the trained imaging settings, demonstrated on in-vivo US data. Our inference runs orders of magnitude faster than optimization-based techniques, enabling future extensions for reconstructing 3D B-mode volumes with only linear computational complexity.

Alvaro Gomariz, Weiye Li, Ece Ozkan et al.

Image-guided radiation therapy can benefit from accurate motion tracking by ultrasound imaging, in order to minimize treatment margins and radiate moving anatomical targets, e.g., due to breathing. One way to formulate this tracking problem is the automatic localization of given tracked anatomical landmarks throughout a temporal ultrasound sequence. For this, we herein propose a fully-convolutional Siamese network that learns the similarity between pairs of image regions containing the same landmark. Accordingly, it learns to localize and thus track arbitrary image features, not only predefined anatomical structures. We employ a temporal consistency model as a location prior, which we combine with the network-predicted location probability map to track a target iteratively in ultrasound sequences. We applied this method on the dataset of the Challenge on Liver Ultrasound Tracking (CLUST) with competitive results, where our work is the first to effectively apply CNNs on this tracking problem, thanks to our temporal regularization.

Firat Ozdemir, Orcun Goksel

Comprehensive surgical planning require complex patient-specific anatomical models. For instance, functional muskuloskeletal simulations necessitate all relevant structures to be segmented, which could be performed in real-time using deep neural networks given sufficient annotated samples. Such large datasets of multiple structure annotations are costly to procure and are often unavailable in practice. Nevertheless, annotations from different studies and centers can be readily available, or become available in the future in an incremental fashion. We propose a class-incremental segmentation framework for extending a deep network trained for some anatomical structure to yet another structure using a small incremental annotation set. Through distilling knowledge from the current state of the framework, we bypass the need for a full retraining. This is a meta-method to extend any choice of desired deep segmentation network with only a minor addition per structure, which makes it suitable for lifelong class-incremental learning and applicable also for future deep neural network architectures. We evaluated our methods on a public knee dataset of 100 MR volumes. Through varying amount of incremental annotation ratios, we show how our proposed method can retain the previous anatomical structure segmentation performance superior to the conventional finetuning approach. In addition, our framework inherently exploits transferable knowledge from previously trained structures to incremental tasks, demonstrated by superior results compared to non-incremental training. With the presented method, new anatomical structures can be learned without catastrophic forgetting of older structures and without extensive increase of memory and complexity.

Lin Zhang, Valery Vishnevskiy, Andras Jakab et al.

Intravoxel incoherent motion (IVIM) imaging allows contrast-agent free in vivo perfusion quantification with magnetic resonance imaging (MRI). However, its use is limited by typically low accuracy due to low signal-to-noise ratio (SNR) at large gradient encoding magnitudes as well as dephasing artefacts caused by subject motion, which is particularly challenging in fetal MRI. To mitigate this problem, we propose an implicit IVIM signal acquisition model with which we learn full posterior distribution of perfusion parameters using artificial neural networks. This posterior then encapsulates the uncertainty of the inferred parameter estimates, which we validate herein via numerical experiments with rejection-based Bayesian sampling. Compared to state-of-the-art IVIM estimation method of segmented least-squares fitting, our proposed approach improves parameter estimation accuracy by 65% on synthetic anisotropic perfusion data. On paired rescans of in vivo fetal MRI, our method increases repeatability of parameter estimation in placenta by 46%.

Valery Vishnevskiy, Sergio J Sanabria, Orcun Goksel

Speed-of-sound is a biomechanical property for quantitative tissue differentiation, with great potential as a new ultrasound-based image modality. A conventional ultrasound array transducer can be used together with an acoustic mirror, or so-called reflector, to reconstruct sound-speed images from time-of-flight measurements to the reflector collected between transducer element pairs, which constitutes a challenging problem of limited-angle computed tomography. For this problem, we herein present a variational network based image reconstruction architecture that is based on optimization loop unrolling, and provide an efficient training protocol of this network architecture on fully synthetic inclusion data. Our results indicate that the learned model presents good generalization ability, being able to reconstruct images with significantly different statistics compared to the training set. Complex inclusion geometries were shown to be successfully reconstructed, also improving over the prior-art by 23% in reconstruction error and by 10% in contrast on synthetic data. In a phantom study, we demonstrated the detection of multiple inclusions that were not distinguishable by prior-art reconstruction, meanwhile improving the contrast by 27% for a stiff inclusion and by 219% for a soft inclusion. Our reconstruction algorithm takes approximately 10ms, enabling its use as a real-time imaging method on an ultrasound machine, for which we are demonstrating an example preliminary setup herein.

Christine Tanner, Firat Ozdemir, Romy Profanter et al.

Deformable Image Registration (DIR) of MR and CT images is one of the most challenging registration task, due to the inherent structural differences of the modalities and the missing dense ground truth. Recently cycle Generative Adversarial Networks (cycle-GANs) have been used to learn the intensity relationship between these 2 modalities for unpaired brain data. Yet its usefulness for DIR was not assessed. In this study we evaluate the DIR performance for thoracic and abdominal organs after synthesis by cycle-GAN. We show that geometric changes, which differentiate the two populations (e.g. inhale vs. exhale), are readily synthesized as well. This causes substantial problems for any application which relies on spatial correspondences being preserved between the real and the synthesized image (e.g. plan, segmentation, landmark propagation). To alleviate this problem, we investigated reducing the spatial information provided to the discriminator by decreasing the size of its receptive fields. Image synthesis was learned from 17 unpaired subjects per modality. Registration performance was evaluated with respect to manual segmentations of 11 structures for 3 subjects from the VISERAL challenge. State-of-the-art DIR methods based on Normalized Mutual Information (NMI), Modality Independent Neighborhood Descriptor (MIND) and their novel combination achieved a mean segmentation overlap ratio of 76.7, 67.7, 76.9%, respectively. This dropped to 69.1% or less when registering images synthesized by cycle-GAN based on local correlation, due to the poor performance on the thoracic region, where large lung volume changes were synthesized. Performance for the abdominal region was similar to that of CT-MRI NMI registration (77.4 vs. 78.8%) when using 3D synthesizing MRIs (12 slices) and medium sized receptive fields for the discriminator.

Firat Ozdemir, Zixuan Peng, Christine Tanner et al.

Segmentation is essential for medical image analysis tasks such as intervention planning, therapy guidance, diagnosis, treatment decisions. Deep learning is becoming increasingly prominent for segmentation, where the lack of annotations, however, often becomes the main limitation. Due to privacy concerns and ethical considerations, most medical datasets are created, curated, and allow access only locally. Furthermore, current deep learning methods are often suboptimal in translating anatomical knowledge between different medical imaging modalities. Active learning can be used to select an informed set of image samples to request for manual annotation, in order to best utilize the limited annotation time of clinical experts for optimal outcomes, which we focus on in this work. Our contributions herein are two fold: (1) we enforce domain-representativeness of selected samples using a proposed penalization scheme to maximize information at the network abstraction layer, and (2) we propose a Borda-count based sample querying scheme for selecting samples for segmentation. Comparative experiments with baseline approaches show that the samples queried with our proposed method, where both above contributions are combined, result in significantly improved segmentation performance for this active learning task.

Firat Ozdemir, Philipp Fuernstahl, Orcun Goksel

Deep learning has been widely accepted as a promising solution for medical image segmentation, given a sufficiently large representative dataset of images with corresponding annotations. With ever increasing amounts of annotated medical datasets, it is infeasible to train a learning method always with all data from scratch. This is also doomed to hit computational limits, e.g., memory or runtime feasible for training. Incremental learning can be a potential solution, where new information (images or anatomy) is introduced iteratively. Nevertheless, for the preservation of the collective information, it is essential to keep some "important" (i.e. representative) images and annotations from the past, while adding new information. In this paper, we introduce a framework for applying incremental learning for segmentation and propose novel methods for selecting representative data therein. We comparatively evaluate our methods in different scenarios using MR images and validate the increased learning capacity with using our methods.

Ece Ozkan, Gemma Roig, Orcun Goksel et al.

We show that the algorithm to extract diverse M -solutions from a Conditional Random Field (called divMbest [1]) takes exactly the form of a Herding procedure [2], i.e. a deterministic dynamical system that produces a sequence of hypotheses that respect a set of observed moment constraints. This generalization enables us to invoke properties of Herding that show that divMbest enforces implausible constraints which may yield wrong assumptions for some problem settings. Our experiments in semantic segmentation demonstrate that seeing divMbest as an instance of Herding leads to better alternatives for the implausible constraints of divMbest.