Joseph Paul Cohen

Reza Azad, Ehsan Khodapanah Aghdam, Amelie Rauland et al.

Automatic medical image segmentation is a crucial topic in the medical domain and successively a critical counterpart in the computer-aided diagnosis paradigm. U-Net is the most widespread image segmentation architecture due to its flexibility, optimized modular design, and success in all medical image modalities. Over the years, the U-Net model achieved tremendous attention from academic and industrial researchers. Several extensions of this network have been proposed to address the scale and complexity created by medical tasks. Addressing the deficiency of the naive U-Net model is the foremost step for vendors to utilize the proper U-Net variant model for their business. Having a compendium of different variants in one place makes it easier for builders to identify the relevant research. Also, for ML researchers it will help them understand the challenges of the biological tasks that challenge the model. To address this, we discuss the practical aspects of the U-Net model and suggest a taxonomy to categorize each network variant. Moreover, to measure the performance of these strategies in a clinical application, we propose fair evaluations of some unique and famous designs on well-known datasets. We provide a comprehensive implementation library with trained models for future research. In addition, for ease of future studies, we created an online list of U-Net papers with their possible official implementation. All information is gathered in https://github.com/NITR098/Awesome-U-Net repository.

Joseph Paul Cohen, Rupert Brooks, Sovann En et al.

This study evaluates the effect of counterfactual explanations on the interpretation of chest X-rays. We conduct a reader study with two radiologists assessing 240 chest X-ray predictions to rate their confidence that the model's prediction is correct using a 5 point scale. Half of the predictions are false positives. Each prediction is explained twice, once using traditional attribution methods and once with a counterfactual explanation. The overall results indicate that counterfactual explanations allow a radiologist to have more confidence in true positive predictions compared to traditional approaches (0.15$\pm$0.95 with p=0.01) with only a small increase in false positive predictions (0.04$\pm$1.06 with p=0.57). We observe the specific prediction tasks of Mass and Atelectasis appear to benefit the most compared to other tasks.

Louis Blankemeier, Ashwin Kumar, Joseph Paul Cohen et al.

The large volume of abdominal computed tomography (CT) scans coupled with the shortage of radiologists have intensified the need for automated medical image analysis tools. Previous state-of-the-art approaches for automated analysis leverage vision-language models (VLMs) that jointly model images and radiology reports. However, current medical VLMs are generally limited to 2D images and short reports. Here to overcome these shortcomings for abdominal CT interpretation, we introduce Merlin, a 3D VLM that learns from volumetric CT scans, electronic health record data and radiology reports. This approach is enabled by a multistage pretraining framework that does not require additional manual annotations. We trained Merlin using a high-quality clinical dataset of paired CT scans (>6 million images from 15,331 CT scans), diagnosis codes (>1.8 million codes) and radiology reports (>6 million tokens). We comprehensively evaluated Merlin on 6 task types and 752 individual tasks that covered diagnostic, prognostic and quality-related tasks. The non-adapted (off-the-shelf) tasks included zero-shot classification of findings (30 findings), phenotype classification (692 phenotypes) and zero-shot cross-modal retrieval (image-to-findings and image-to-impression). The model-adapted tasks included 5-year chronic disease prediction (6 diseases), radiology report generation and 3D semantic segmentation (20 organs). We validated Merlin at scale, with internal testing on 5,137 CT scans and external testing on 44,098 CT scans from 3 independent sites and 2 public datasets. The results demonstrated high generalization across institutions and anatomies. Merlin outperformed 2D VLMs, CT foundation models and off-the-shelf radiology models. We also release our trained models, code, and dataset, available at: https://github.com/StanfordMIMI/Merlin.

Arjun Soin, Jameson Merkow, Jin Long et al.

Clinical Artificial lntelligence (AI) applications are rapidly expanding worldwide, and have the potential to impact to all areas of medical practice. Medical imaging applications constitute a vast majority of approved clinical AI applications. Though healthcare systems are eager to adopt AI solutions a fundamental question remains: \textit{what happens after the AI model goes into production?} We use the CheXpert and PadChest public datasets to build and test a medical imaging AI drift monitoring workflow to track data and model drift without contemporaneous ground truth. We simulate drift in multiple experiments to compare model performance with our novel multi-modal drift metric, which uses DICOM metadata, image appearance representation from a variational autoencoder (VAE), and model output probabilities as input. Through experimentation, we demonstrate a strong proxy for ground truth performance using unsupervised distributional shifts in relevant metadata, predicted probabilities, and VAE latent representation. Our key contributions include (1) proof-of-concept for medical imaging drift detection that includes the use of VAE and domain specific statistical methods, (2) a multi-modal methodology to measure and unify drift metrics, (3) new insights into the challenges and solutions to observe deployed medical imaging AI, and (4) creation of open-source tools that enable others to easily run their own workflows and scenarios. This work has important implications. It addresses the concerning translation gap found in continuous medical imaging AI model monitoring common in dynamic healthcare environments.

Joseph Paul Cohen, Joseph D. Viviano, Paul Bertin et al.

TorchXRayVision is an open source software library for working with chest X-ray datasets and deep learning models. It provides a common interface and common pre-processing chain for a wide set of publicly available chest X-ray datasets. In addition, a number of classification and representation learning models with different architectures, trained on different data combinations, are available through the library to serve as baselines or feature extractors.

Andreanne Lemay, Charley Gros, Olivier Vincent et al.

Medical images are often accompanied by metadata describing the image (vendor, acquisition parameters) and the patient (disease type or severity, demographics, genomics). This metadata is usually disregarded by image segmentation methods. In this work, we adapt a linear conditioning method called FiLM (Feature-wise Linear Modulation) for image segmentation tasks. This FiLM adaptation enables integrating metadata into segmentation models for better performance. We observed an average Dice score increase of 5.1% on spinal cord tumor segmentation when incorporating the tumor type with FiLM. The metadata modulates the segmentation process through low-cost affine transformations applied on feature maps which can be included in any neural network's architecture. Additionally, we assess the relevance of segmentation FiLM layers for tackling common challenges in medical imaging: multi-class training with missing segmentations, model adaptation to multiple tasks, and training with a limited or unbalanced number of annotated data. Our results demonstrated the following benefits of FiLM for segmentation: FiLMed U-Net was robust to missing labels and reached higher Dice scores with few labels (up to 16.7%) compared to single-task U-Net. The code is open-source and available at www.ivadomed.org.

Joseph Paul Cohen, Rupert Brooks, Sovann En et al.

Motivation: Traditional image attribution methods struggle to satisfactorily explain predictions of neural networks. Prediction explanation is important, especially in medical imaging, for avoiding the unintended consequences of deploying AI systems when false positive predictions can impact patient care. Thus, there is a pressing need to develop improved models for model explainability and introspection. Specific problem: A new approach is to transform input images to increase or decrease features which cause the prediction. However, current approaches are difficult to implement as they are monolithic or rely on GANs. These hurdles prevent wide adoption. Our approach: Given an arbitrary classifier, we propose a simple autoencoder and gradient update (Latent Shift) that can transform the latent representation of a specific input image to exaggerate or curtail the features used for prediction. We use this method to study chest X-ray classifiers and evaluate their performance. We conduct a reader study with two radiologists assessing 240 chest X-ray predictions to identify which ones are false positives (half are) using traditional attribution maps or our proposed method. Results: We found low overlap with ground truth pathology masks for models with reasonably high accuracy. However, the results from our reader study indicate that these models are generally looking at the correct features. We also found that the Latent Shift explanation allows a user to have more confidence in true positive predictions compared to traditional approaches (0.15$\pm$0.95 in a 5 point scale with p=0.01) with only a small increase in false positive predictions (0.04$\pm$1.06 with p=0.57). Accompanying webpage: https://mlmed.org/gifsplanation Source code: https://github.com/mlmed/gifsplanation

Charley Gros, Andreanne Lemay, Olivier Vincent et al.

ivadomed is an open-source Python package for designing, end-to-end training, and evaluating deep learning models applied to medical imaging data. The package includes APIs, command-line tools, documentation, and tutorials. ivadomed also includes pre-trained models such as spinal tumor segmentation and vertebral labeling. Original features of ivadomed include a data loader that can parse image metadata (e.g., acquisition parameters, image contrast, resolution) and subject metadata (e.g., pathology, age, sex) for custom data splitting or extra information during training and evaluation. Any dataset following the Brain Imaging Data Structure (BIDS) convention will be compatible with ivadomed without the need to manually organize the data, which is typically a tedious task. Beyond the traditional deep learning methods, ivadomed features cutting-edge architectures, such as FiLM and HeMis, as well as various uncertainty estimation methods (aleatoric and epistemic), and losses adapted to imbalanced classes and non-binary predictions. Each step is conveniently configurable via a single file. At the same time, the code is highly modular to allow addition/modification of an architecture or pre/post-processing steps. Example applications of ivadomed include MRI object detection, segmentation, and labeling of anatomical and pathological structures. Overall, ivadomed enables easy and quick exploration of the latest advances in deep learning for medical imaging applications. ivadomed's main project page is available at https://ivadomed.org.

Karsten Roth, Timo Milbich, Björn Ommer et al.

Deep Metric Learning (DML) provides a crucial tool for visual similarity and zero-shot applications by learning generalizing embedding spaces, although recent work in DML has shown strong performance saturation across training objectives. However, generalization capacity is known to scale with the embedding space dimensionality. Unfortunately, high dimensional embeddings also create higher retrieval cost for downstream applications. To remedy this, we propose \emph{Simultaneous Similarity-based Self-distillation (S2SD). S2SD extends DML with knowledge distillation from auxiliary, high-dimensional embedding and feature spaces to leverage complementary context during training while retaining test-time cost and with negligible changes to the training time. Experiments and ablations across different objectives and standard benchmarks show S2SD offers notable improvements of up to 7% in Recall@1, while also setting a new state-of-the-art. Code available at https://github.com/MLforHealth/S2SD.

Hasib Zunair, Aimon Rahman, Nabeel Mohammed et al.

A common approach to medical image analysis on volumetric data uses deep 2D convolutional neural networks (CNNs). This is largely attributed to the challenges imposed by the nature of the 3D data: variable volume size, GPU exhaustion during optimization. However, dealing with the individual slices independently in 2D CNNs deliberately discards the depth information which results in poor performance for the intended task. Therefore, it is important to develop methods that not only overcome the heavy memory and computation requirements but also leverage the 3D information. To this end, we evaluate a set of volume uniformizing methods to address the aforementioned issues. The first method involves sampling information evenly from a subset of the volume. Another method exploits the full geometry of the 3D volume by interpolating over the z-axis. We demonstrate performance improvements using controlled ablation studies as well as put this approach to the test on the ImageCLEF Tuberculosis Severity Assessment 2019 benchmark. We report 73% area under curve (AUC) and binary classification accuracy (ACC) of 67.5% on the test set beating all methods which leveraged only image information (without using clinical meta-data) achieving 5-th position overall. All codes and models are made available at https://github.com/hasibzunair/uniformizing-3D.

Joseph Paul Cohen, Paul Morrison, Lan Dao et al.

Across the world's coronavirus disease 2019 (COVID-19) hot spots, the need to streamline patient diagnosis and management has become more pressing than ever. As one of the main imaging tools, chest X-rays (CXRs) are common, fast, non-invasive, relatively cheap, and potentially bedside to monitor the progression of the disease. This paper describes the first public COVID-19 image data collection as well as a preliminary exploration of possible use cases for the data. This dataset currently contains hundreds of frontal view X-rays and is the largest public resource for COVID-19 image and prognostic data, making it a necessary resource to develop and evaluate tools to aid in the treatment of COVID-19. It was manually aggregated from publication figures as well as various web based repositories into a machine learning (ML) friendly format with accompanying dataloader code. We collected frontal and lateral view imagery and metadata such as the time since first symptoms, intensive care unit (ICU) status, survival status, intubation status, or hospital location. We present multiple possible use cases for the data such as predicting the need for the ICU, predicting patient survival, and understanding a patient's trajectory during treatment. Data can be accessed here: https://github.com/ieee8023/covid-chestxray-dataset

Joseph Paul Cohen, Lan Dao, Paul Morrison et al.

Purpose: The need to streamline patient management for COVID-19 has become more pressing than ever. Chest X-rays provide a non-invasive (potentially bedside) tool to monitor the progression of the disease. In this study, we present a severity score prediction model for COVID-19 pneumonia for frontal chest X-ray images. Such a tool can gauge severity of COVID-19 lung infections (and pneumonia in general) that can be used for escalation or de-escalation of care as well as monitoring treatment efficacy, especially in the ICU. Methods: Images from a public COVID-19 database were scored retrospectively by three blinded experts in terms of the extent of lung involvement as well as the degree of opacity. A neural network model that was pre-trained on large (non-COVID-19) chest X-ray datasets is used to construct features for COVID-19 images which are predictive for our task. Results: This study finds that training a regression model on a subset of the outputs from an this pre-trained chest X-ray model predicts our geographic extent score (range 0-8) with 1.14 mean absolute error (MAE) and our lung opacity score (range 0-6) with 0.78 MAE. Conclusions: These results indicate that our model's ability to gauge severity of COVID-19 lung infections could be used for escalation or de-escalation of care as well as monitoring treatment efficacy, especially in the intensive care unit (ICU). A proper clinical trial is needed to evaluate efficacy. To enable this we make our code, labels, and data available online at https://github.com/mlmed/torchxrayvision/tree/master/scripts/covid-severity and https://github.com/ieee8023/covid-chestxray-dataset

Lucas Rouhier, Francisco Perdigon Romero, Joseph Paul Cohen et al.

Labeling intervertebral discs is relevant as it notably enables clinicians to understand the relationship between a patient's symptoms (pain, paralysis) and the exact level of spinal cord injury. However manually labeling those discs is a tedious and user-biased task which would benefit from automated methods. While some automated methods already exist for MRI and CT-scan, they are either not publicly available, or fail to generalize across various imaging contrasts. In this paper we combine a Fully Convolutional Network (FCN) with inception modules to localize and label intervertebral discs. We demonstrate a proof-of-concept application in a publicly-available multi-center and multi-contrast MRI database (n=235 subjects). The code is publicly available at https://github.com/neuropoly/vertebral-labeling-deep-learning.

Karsten Roth, Timo Milbich, Samarth Sinha et al.

Deep Metric Learning (DML) is arguably one of the most influential lines of research for learning visual similarities with many proposed approaches every year. Although the field benefits from the rapid progress, the divergence in training protocols, architectures, and parameter choices make an unbiased comparison difficult. To provide a consistent reference point, we revisit the most widely used DML objective functions and conduct a study of the crucial parameter choices as well as the commonly neglected mini-batch sampling process. Under consistent comparison, DML objectives show much higher saturation than indicated by literature. Further based on our analysis, we uncover a correlation between the embedding space density and compression to the generalization performance of DML models. Exploiting these insights, we propose a simple, yet effective, training regularization to reliably boost the performance of ranking-based DML models on various standard benchmark datasets. Code and a publicly accessible WandB-repo are available at https://github.com/Confusezius/Revisiting_Deep_Metric_Learning_PyTorch.

Joseph Paul Cohen, Mohammad Hashir, Rupert Brooks et al.

This large scale study focuses on quantifying what X-rays diagnostic prediction tasks generalize well across multiple different datasets. We present evidence that the issue of generalization is not due to a shift in the images but instead a shift in the labels. We study the cross-domain performance, agreement between models, and model representations. We find interesting discrepancies between performance and agreement where models which both achieve good performance disagree in their predictions as well as models which agree yet achieve poor performance. We also test for concept similarity by regularizing a network to group tasks across multiple datasets together and observe variation across the tasks. All code is made available online and data is publicly available: https://github.com/mlmed/torchxrayvision

Tristan Deleu, Tobias Würfl, Mandana Samiei et al.

The constant introduction of standardized benchmarks in the literature has helped accelerating the recent advances in meta-learning research. They offer a way to get a fair comparison between different algorithms, and the wide range of datasets available allows full control over the complexity of this evaluation. However, for a large majority of code available online, the data pipeline is often specific to one dataset, and testing on another dataset requires significant rework. We introduce Torchmeta, a library built on top of PyTorch that enables seamless and consistent evaluation of meta-learning algorithms on multiple datasets, by providing data-loaders for most of the standard benchmarks in few-shot classification and regression, with a new meta-dataset abstraction. It also features some extensions for PyTorch to simplify the development of models compatible with meta-learning algorithms. The code is available here: https://github.com/tristandeleu/pytorch-meta

Joseph Paul Cohen, Paul Bertin, Vincent Frappier

In order to bridge the gap between Deep Learning researchers and medical professionals we develop a very accessible free prototype system which can be used by medical professionals to understand the reality of Deep Learning tools for chest X-ray diagnostics. The system is designed to be a second opinion where a user can process an image to confirm or aid in their diagnosis. Code and network weights are delivered via a URL to a web browser (including cell phones) but the patient data remains on the users machine and all processing occurs locally. This paper discusses the three main components in detail: out-of-distribution detection, disease prediction, and prediction explanation. The system open source and freely available here: https://mlmed.org/tools/xray

Zhihong Chen, Maya Varma, Justin Xu et al. · mila, oxford

Over 1.4 billion chest X-rays (CXRs) are performed annually due to their cost-effectiveness as an initial diagnostic test. This scale of radiological studies provides a significant opportunity to streamline CXR interpretation and documentation. While foundation models are a promising solution, the lack of publicly available large-scale datasets and benchmarks inhibits their iterative development and real-world evaluation. To overcome these challenges, we constructed a large-scale dataset (CheXinstruct), which we utilized to train a vision-language foundation model (CheXagent). We systematically demonstrated competitive performance across eight distinct task types on our novel evaluation benchmark (CheXbench). Beyond technical validation, we assessed the real-world utility of CheXagent in directly drafting radiology reports. Our clinical assessment with eight radiologists revealed a 36% time saving for residents using CheXagent-drafted reports, while attending radiologists showed no significant time difference editing resident-drafted or CheXagent-drafted reports. The CheXagent-drafted reports improved the writing efficiency of both radiology residents and attending radiologists in 81% and 61% of cases, respectively, without loss of quality. Overall, we demonstrate that CheXagent can effectively perform a variety of CXR interpretation tasks and holds potential to assist radiologists in routine clinical workflows.

Suhas BN, Han-Chin Shing, Lei Xu et al.

Hallucinations in large language models (LLMs) during summarization of patient-clinician dialogues pose significant risks to patient care and clinical decision-making. However, the phenomenon remains understudied in the clinical domain, with uncertainty surrounding the applicability of general-domain hallucination detectors. The rarity and randomness of hallucinations further complicate their investigation. In this paper, we conduct an evaluation of hallucination detection methods in the medical domain, and construct two datasets for the purpose: A fact-controlled Leave-N-out dataset -- generated by systematically removing facts from source dialogues to induce hallucinated content in summaries; and a natural hallucination dataset -- arising organically during LLM-based medical summarization. We show that general-domain detectors struggle to detect clinical hallucinations, and that performance on fact-controlled hallucinations does not reliably predict effectiveness on natural hallucinations. We then develop fact-based approaches that count hallucinations, offering explainability not available with existing methods. Notably, our LLM-based detectors, which we developed using fact-controlled hallucinations, generalize well to detecting real-world clinical hallucinations. This research contributes a suite of specialized metrics supported by expert-annotated datasets to advance faithful clinical summarization systems.

Joseph Paul Cohen, Louis Blankemeier, Akshay Chaudhari

Counterfactual explanations enhance the interpretability of deep learning models in medical imaging, yet adapting them to 3D CT scans poses challenges due to volumetric complexity and resource demands. We extend the Latent Shift counterfactual generation method from 2D applications to explain 3D computed tomography (CT) scans classifiers. We address the challenges associated with 3D classifiers, such as limited training samples and high memory demands, by implementing a slice-based autoencoder and gradient blocking except for specific chunks of slices. This method leverages a 2D encoder trained on CT slices, which are subsequently combined to maintain 3D context. We demonstrate this technique on two models for clinical phenotype prediction and lung segmentation. Our approach is both memory-efficient and effective for generating interpretable counterfactuals in high-resolution 3D medical imaging.

Enoch Tetteh, Joseph Viviano, Yoshua Bengio et al.

Learning models that generalize under different distribution shifts in medical imaging has been a long-standing research challenge. There have been several proposals for efficient and robust visual representation learning among vision research practitioners, especially in the sensitive and critical biomedical domain. In this paper, we propose an idea for out-of-distribution generalization of chest X-ray pathologies that uses a simple balanced batch sampling technique. We observed that balanced sampling between the multiple training datasets improves the performance over baseline models trained without balancing.

Tianshi Cao, Chin-Wei Huang, David Yu-Tung Hui et al.

Motivation: Deep learning models deployed for use on medical tasks can be equipped with Out-of-Distribution Detection (OoDD) methods in order to avoid erroneous predictions. However it is unclear which OoDD method should be used in practice. Specific Problem: Systems trained for one particular domain of images cannot be expected to perform accurately on images of a different domain. These images should be flagged by an OoDD method prior to diagnosis. Our approach: This paper defines 3 categories of OoD examples and benchmarks popular OoDD methods in three domains of medical imaging: chest X-ray, fundus imaging, and histology slides. Results: Our experiments show that despite methods yielding good results on some categories of out-of-distribution samples, they fail to recognize images close to the training distribution. Conclusion: We find a simple binary classifier on the feature representation has the best accuracy and AUPRC on average. Users of diagnostic tools which employ these OoDD methods should still remain vigilant that images very close to the training distribution yet not in it could yield unexpected results.

Timo Milbich, Karsten Roth, Homanga Bharadhwaj et al.

Visual Similarity plays an important role in many computer vision applications. Deep metric learning (DML) is a powerful framework for learning such similarities which not only generalize from training data to identically distributed test distributions, but in particular also translate to unknown test classes. However, its prevailing learning paradigm is class-discriminative supervised training, which typically results in representations specialized in separating training classes. For effective generalization, however, such an image representation needs to capture a diverse range of data characteristics. To this end, we propose and study multiple complementary learning tasks, targeting conceptually different data relationships by only resorting to the available training samples and labels of a standard DML setting. Through simultaneous optimization of our tasks we learn a single model to aggregate their training signals, resulting in strong generalization and state-of-the-art performance on multiple established DML benchmark datasets.

Joseph Paul Cohen, Paul Morrison, Lan Dao

This paper describes the initial COVID-19 open image data collection. It was created by assembling medical images from websites and publications and currently contains 123 frontal view X-rays.

Olivier Vincent, Charley Gros, Joseph Paul Cohen et al.

Despite recent improvements in medical image segmentation, the ability to generalize across imaging contrasts remains an open issue. To tackle this challenge, we implement Feature-wise Linear Modulation (FiLM) to leverage physics knowledge within the segmentation model and learn the characteristics of each contrast. Interestingly, a well-optimised U-Net reached the same performance as our FiLMed-Unet on a multi-contrast dataset (0.72 of Dice score), which suggests that there is a bottleneck in spinal MS lesion segmentation different from the generalization across varying contrasts. This bottleneck likely stems from inter-rater variability, which is estimated at 0.61 of Dice score in our dataset.

Mohammad Hashir, Hadrien Bertrand, Joseph Paul Cohen

Most deep learning models in chest X-ray prediction utilize the posteroanterior (PA) view due to the lack of other views available. PadChest is a large-scale chest X-ray dataset that has almost 200 labels and multiple views available. In this work, we use PadChest to explore multiple approaches to merging the PA and lateral views for predicting the radiological labels associated with the X-ray image. We find that different methods of merging the model utilize the lateral view differently. We also find that including the lateral view increases performance for 32 labels in the dataset, while being neutral for the others. The increase in overall performance is comparable to the one obtained by using only the PA view with twice the amount of patients in the training set.

Mohammad Hashir, Paul Bertin, Martin Weiss et al.

Gene interaction graphs aim to capture various relationships between genes and represent decades of biology research. When trying to make predictions from genomic data, those graphs could be used to overcome the curse of dimensionality by making machine learning models sparser and more consistent with biological common knowledge. In this work, we focus on assessing whether those graphs capture dependencies seen in gene expression data better than random. We formulate a condition that graphs should satisfy to provide a good prior knowledge and propose to test it using a `Single Gene Inference' (SGI) task. We compare random graphs with seven major gene interaction graphs published by different research groups, aiming to measure the true benefit of using biologically relevant graphs in this context. Our analysis finds that dependencies can be captured almost as well at random which suggests that, in terms of gene expression levels, the relevant information about the state of the cell is spread across many genes.

Shawn Tan, Guillaume Androz, Ahmad Chamseddine et al.

We release the largest public ECG dataset of continuous raw signals for representation learning containing 11 thousand patients and 2 billion labelled beats. Our goal is to enable semi-supervised ECG models to be made as well as to discover unknown subtypes of arrhythmia and anomalous ECG signal events. To this end, we propose an unsupervised representation learning task, evaluated in a semi-supervised fashion. We provide a set of baselines for different feature extractors that can be built upon. Additionally, we perform qualitative evaluations on results from PCA embeddings, where we identify some clustering of known subtypes indicating the potential for representation learning in arrhythmia sub-type discovery.

Mandana Samiei, Tobias Würfl, Tristan Deleu et al.

Machine learning is bringing a paradigm shift to healthcare by changing the process of disease diagnosis and prognosis in clinics and hospitals. This development equips doctors and medical staff with tools to evaluate their hypotheses and hence make more precise decisions. Although most current research in the literature seeks to develop techniques and methods for predicting one particular clinical outcome, this approach is far from the reality of clinical decision making in which you have to consider several factors simultaneously. In addition, it is difficult to follow the recent progress concretely as there is a lack of consistency in benchmark datasets and task definitions in the field of Genomics. To address the aforementioned issues, we provide a clinical Meta-Dataset derived from the publicly available data hub called The Cancer Genome Atlas Program (TCGA) that contains 174 tasks. We believe those tasks could be good proxy tasks to develop methods which can work on a few samples of gene expression data. Also, learning to predict multiple clinical variables using gene-expression data is an important task due to the variety of phenotypes in clinical problems and lack of samples for some of the rare variables. The defined tasks cover a wide range of clinical problems including predicting tumor tissue site, white cell count, histological type, family history of cancer, gender, and many others which we explain later in the paper. Each task represents an independent dataset. We use regression and neural network baselines for all the tasks using only 150 samples and compare their performance.

Saeid Asgari Taghanaki, Kumar Abhishek, Joseph Paul Cohen et al.

The semantic image segmentation task consists of classifying each pixel of an image into an instance, where each instance corresponds to a class. This task is a part of the concept of scene understanding or better explaining the global context of an image. In the medical image analysis domain, image segmentation can be used for image-guided interventions, radiotherapy, or improved radiological diagnostics. In this review, we categorize the leading deep learning-based medical and non-medical image segmentation solutions into six main groups of deep architectural, data synthesis-based, loss function-based, sequenced models, weakly supervised, and multi-task methods and provide a comprehensive review of the contributions in each of these groups. Further, for each group, we analyze each variant of these groups and discuss the limitations of the current approaches and present potential future research directions for semantic image segmentation.

Joseph D. Viviano, Becks Simpson, Francis Dutil et al.

Poor generalization is one symptom of models that learn to predict target variables using spuriously-correlated image features present only in the training distribution instead of the true image features that denote a class. It is often thought that this can be diagnosed visually using attribution (aka saliency) maps. We study if this assumption is correct. In some prediction tasks, such as for medical images, one may have some images with masks drawn by a human expert, indicating a region of the image containing relevant information to make the prediction. We study multiple methods that take advantage of such auxiliary labels, by training networks to ignore distracting features which may be found outside of the region of interest. This mask information is only used during training and has an impact on generalization accuracy depending on the severity of the shift between the training and test distributions. Surprisingly, while these methods improve generalization performance in the presence of a covariate shift, there is no strong correspondence between the correction of attribution towards the features a human expert has labelled as important and generalization performance. These results suggest that the root cause of poor generalization may not always be spatially defined, and raise questions about the utility of masks as "attribution priors" as well as saliency maps for explainable predictions.

Paul Bertin, Mohammad Hashir, Martin Weiss et al.

Gene interaction graphs aim to capture various relationships between genes and can represent decades of biology research. When trying to make predictions from genomic data, those graphs could be used to overcome the curse of dimensionality by making machine learning models sparser and more consistent with biological common knowledge. In this work, we focus on assessing how well those graphs capture dependencies seen in gene expression data to evaluate the adequacy of the prior knowledge provided by those graphs. We propose a condition graphs should satisfy to provide good prior knowledge and test it using `Single Gene Inference' tasks. We also compare with randomly generated graphs, aiming to measure the true benefit of using biologically relevant graphs in this context, and validate our findings with five clinical tasks. We find some graphs capture relevant dependencies for most genes while being very sparse. Our analysis with random graphs finds that dependencies can be captured almost as well at random which suggests that, in terms of gene expression levels, the relevant information about the state of the cell is spread across many genes.

Hadrien Bertrand, Mohammad Hashir, Joseph Paul Cohen

Most convolutional neural networks in chest radiology use only the frontal posteroanterior (PA) view to make a prediction. However the lateral view is known to help the diagnosis of certain diseases and conditions. The recently released PadChest dataset contains paired PA and lateral views, allowing us to study for which diseases and conditions the performance of a neural network improves when provided a lateral x-ray view as opposed to a frontal posteroanterior (PA) view. Using a simple DenseNet model, we find that using the lateral view increases the AUC of 8 of the 56 labels in our data and achieves the same performance as the PA view for 21 of the labels. We find that using the PA and lateral views jointly doesn't trivially lead to an increase in performance but suggest further investigation.

Becks Simpson, Francis Dutil, Yoshua Bengio et al.

With too few samples or too many model parameters, overfitting can inhibit the ability to generalise predictions to new data. Within medical imaging, this can occur when features are incorrectly assigned importance such as distinct hospital specific artifacts, leading to poor performance on a new dataset from a different institution without those features, which is undesirable. Most regularization methods do not explicitly penalize the incorrect association of these features to the target class and hence fail to address this issue. We propose a regularization method, GradMask, which penalizes saliency maps inferred from the classifier gradients when they are not consistent with the lesion segmentation. This prevents non-tumor related features to contribute to the classification of unhealthy samples. We demonstrate that this method can improve test accuracy between 1-3% compared to the baseline without GradMask, showing that it has an impact on reducing overfitting.

Martin Weiss, Margaux Luck, Roger Girgis et al.

The number of visually impaired or blind (VIB) people in the world is estimated at several hundred million. Based on a series of interviews with the VIB and developers of assistive technology, this paper provides a survey of machine-learning based mobile applications and identifies the most relevant applications. We discuss the functionality of these apps, how they align with the needs and requirements of the VIB users, and how they can be improved with techniques such as federated learning and model compression. As a result of this study we identify promising future directions of research in mobile perception, micro-navigation, and content-summarization.

Assya Trofimov, Francis Dutil, Claude Perreault et al.

In this work we propose a method to compute continuous embeddings for kmers from raw RNA-seq data, without the need for alignment to a reference genome. The approach uses an RNN to transform kmers of the RNA-seq reads into a 2 dimensional representation that is used to predict abundance of each kmer. We report that our model captures information of both DNA sequence similarity as well as DNA sequence abundance in the embedding latent space, that we call the Latent Transcriptome. We confirm the quality of these vectors by comparing them to known gene sub-structures and report that the latent space recovers exon information from raw RNA-Seq data from acute myeloid leukemia patients. Furthermore we show that this latent space allows the detection of genomic abnormalities such as translocations as well as patient-specific mutations, making this representation space both useful for visualization as well as analysis.

Francis Dutil, Joseph Paul Cohen, Martin Weiss et al.

We study the challenges of applying deep learning to gene expression data. We find experimentally that there exists non-linear signal in the data, however is it not discovered automatically given the noise and low numbers of samples used in most research. We discuss how gene interaction graphs (same pathway, protein-protein, co-expression, or research paper text association) can be used to impose a bias on a deep model similar to the spatial bias imposed by convolutions on an image. We explore the usage of Graph Convolutional Neural Networks coupled with dropout and gene embeddings to utilize the graph information. We find this approach provides an advantage for particular tasks in a low data regime but is very dependent on the quality of the graph used. We conclude that more work should be done in this direction. We design experiments that show why existing methods fail to capture signal that is present in the data when features are added which clearly isolates the problem that needs to be addressed.

Margaux Luck, Tristan Sylvain, Joseph Paul Cohen et al.

Survival analysis is a type of semi-supervised ranking task where the target output (the survival time) is often right-censored. Utilizing this information is a challenge because it is not obvious how to correctly incorporate these censored examples into a model. We study how three categories of loss functions, namely partial likelihood methods, rank methods, and our classification method based on a Wasserstein metric (WM) and the non-parametric Kaplan Meier estimate of the probability density to impute the labels of censored examples, can take advantage of this information. The proposed method allows us to have a model that predict the probability distribution of an event. If a clinician had access to the detailed probability of an event over time this would help in treatment planning. For example, determining if the risk of kidney graft rejection is constant or peaked after some time. Also, we demonstrate that this approach directly optimizes the expected C-index which is the most common evaluation metric for ranking survival models.

Joseph Paul Cohen, Margaux Luck, Sina Honari

This paper discusses how distribution matching losses, such as those used in CycleGAN, when used to synthesize medical images can lead to mis-diagnosis of medical conditions. It seems appealing to use these new image synthesis methods for translating images from a source to a target domain because they can produce high quality images and some even do not require paired data. However, the basis of how these image translation models work is through matching the translation output to the distribution of the target domain. This can cause an issue when the data provided in the target domain has an over or under representation of some classes (e.g. healthy or sick). When the output of an algorithm is a transformed image there are uncertainties whether all known and unknown class labels have been preserved or changed. Therefore, we recommend that these translated images should not be used for direct interpretation (e.g. by doctors) because they may lead to misdiagnosis of patients based on hallucinated image features by an algorithm that matches a distribution. However there are many recent papers that seem as though this is the goal.

Alex Lamb, Devon Hjelm, Yaroslav Ganin et al.

Directed latent variable models that formulate the joint distribution as $p(x,z) = p(z) p(x \mid z)$ have the advantage of fast and exact sampling. However, these models have the weakness of needing to specify $p(z)$, often with a simple fixed prior that limits the expressiveness of the model. Undirected latent variable models discard the requirement that $p(z)$ be specified with a prior, yet sampling from them generally requires an iterative procedure such as blocked Gibbs-sampling that may require many steps to draw samples from the joint distribution $p(x, z)$. We propose a novel approach to learning the joint distribution between the data and a latent code which uses an adversarially learned iterative procedure to gradually refine the joint distribution, $p(x, z)$, to better match with the data distribution on each step. GibbsNet is the best of both worlds both in theory and in practice. Achieving the speed and simplicity of a directed latent variable model, it is guaranteed (assuming the adversarial game reaches the virtual training criteria global minimum) to produce samples from $p(x, z)$ with only a few sampling iterations. Achieving the expressiveness and flexibility of an undirected latent variable model, GibbsNet does away with the need for an explicit $p(z)$ and has the ability to do attribute prediction, class-conditional generation, and joint image-attribute modeling in a single model which is not trained for any of these specific tasks. We show empirically that GibbsNet is able to learn a more complex $p(z)$ and show that this leads to improved inpainting and iterative refinement of $p(x, z)$ for dozens of steps and stable generation without collapse for thousands of steps, despite being trained on only a few steps.

Joseph Paul Cohen, Genevieve Boucher, Craig A. Glastonbury et al.

Counting objects in digital images is a process that should be replaced by machines. This tedious task is time consuming and prone to errors due to fatigue of human annotators. The goal is to have a system that takes as input an image and returns a count of the objects inside and justification for the prediction in the form of object localization. We repose a problem, originally posed by Lempitsky and Zisserman, to instead predict a count map which contains redundant counts based on the receptive field of a smaller regression network. The regression network predicts a count of the objects that exist inside this frame. By processing the image in a fully convolutional way each pixel is going to be accounted for some number of times, the number of windows which include it, which is the size of each window, (i.e., 32x32 = 1024). To recover the true count we take the average over the redundant predictions. Our contribution is redundant counting instead of predicting a density map in order to average over errors. We also propose a novel deep neural network architecture adapted from the Inception family of networks called the Count-ception network. Together our approach results in a 20% relative improvement (2.9 to 2.3 MAE) over the state of the art method by Xie, Noble, and Zisserman in 2016.

Joseph Paul Cohen, Wei Ding, Caitlin Kuhlman et al.

This work describes algorithms for performing discrete object detection, specifically in the case of buildings, where usually only low quality RGB-only geospatial reflective imagery is available. We utilize new candidate search and feature extraction techniques to reduce the problem to a machine learning (ML) classification task. Here we can harness the complex patterns of contrast features contained in training data to establish a model of buildings. We avoid costly sliding windows to generate candidates; instead we innovatively stitch together well known image processing techniques to produce candidates for building detection that cover 80-85% of buildings. Reducing the number of possible candidates is important due to the scale of the problem. Each candidate is subjected to classification which, although linear, costs time and prohibits large scale evaluation. We propose a candidate alignment algorithm to boost classification performance to 80-90% precision with a linear time algorithm and show it has negligible cost. Also, we propose a new concept called a Permutable Haar Mesh (PHM) which we use to form and traverse a search space to recover candidate buildings which were lost in the initial preprocessing phase.

Joseph Paul Cohen, Henry Z. Lo, Wei Ding

Filters in convolutional neural networks are sensitive to their initialization. The random numbers used to initialize filters are a bias and determine if you will "win" and converge to a satisfactory local minimum so we call this The Filter Lottery. We observe that the 28x28 Inception-V3 model without Batch Normalization fails to train 26% of the time when varying the random seed alone. This is a problem that affects the trial and error process of designing a network. Because random seeds have a large impact it makes it hard to evaluate a network design without trying many different random starting weights. This work aims to reduce the bias imposed by the initial weights so a network converges more consistently. We propose to evaluate and replace specific convolutional filters that have little impact on the prediction. We use the gradient norm to evaluate the impact of a filter on error, and re-initialize filters when the gradient norm of its weights falls below a specific threshold. This consistently improves accuracy on the 28x28 Inception-V3 with a median increase of +3.3%. In effect our method RandomOut increases the number of filters explored without increasing the size of the network. We observe that the RandomOut method has more consistent generalization performance, having a standard deviation of 1.3% instead of 2% when varying random seeds, and does so faster and with fewer parameters.

Joseph Paul Cohen, Henry Z. Lo, Tingting Lu et al.

Craters are among the most studied geomorphic features in the Solar System because they yield important information about the past and present geological processes and provide information about the relative ages of observed geologic formations. We present a method for automatic crater detection using advanced machine learning to deal with the large amount of satellite imagery collected. The challenge of automatically detecting craters comes from their is complex surface because their shape erodes over time to blend into the surface. Bandeira provided a seminal dataset that embodied this challenge that is still an unsolved pattern recognition problem to this day. There has been work to solve this challenge based on extracting shape and contrast features and then applying classification models on those features. The limiting factor in this existing work is the use of hand crafted filters on the image such as Gabor or Sobel filters or Haar features. These hand crafted methods rely on domain knowledge to construct. We would like to learn the optimal filters and features based on training examples. In order to dynamically learn filters and features we look to Convolutional Neural Networks (CNNs) which have shown their dominance in computer vision. The power of CNNs is that they can learn image filters which generate features for high accuracy classification.

Joseph Paul Cohen, Wei Ding, Abraham Bagherjeiran

We discuss a key problem in information extraction which deals with wrapper failures due to changing content templates. A good proportion of wrapper failures are due to HTML templates changing to cause wrappers to become incompatible after element inclusion or removal in a DOM (Tree representation of HTML). We perform a large-scale empirical analyses of the causes of shift and mathematically quantify the levels of domain difficulty based on entropy. We propose the XTreePath annotation method to captures contextual node information from the training DOM. We then utilize this annotation in a supervised manner at test time with our proposed Recursive Tree Matching method which locates nodes most similar in context recursively using the tree edit distance. The search is based on a heuristic function that takes into account the similarity of a tree compared to the structure that was present in the training data. We evaluate XTreePath using 117,422 pages from 75 diverse websites in 8 vertical markets. Our XTreePath method consistently outperforms XPath and a current commercial system in terms of successful extractions in a blackbox test. We make our code and datasets publicly available online.