Daniel L. Rubin

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

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

Rui Yan, Liangqiong Qu, Qingyue Wei et al.

The collection and curation of large-scale medical datasets from multiple institutions is essential for training accurate deep learning models, but privacy concerns often hinder data sharing. Federated learning (FL) is a promising solution that enables privacy-preserving collaborative learning among different institutions, but it generally suffers from performance deterioration due to heterogeneous data distributions and a lack of quality labeled data. In this paper, we present a robust and label-efficient self-supervised FL framework for medical image analysis. Our method introduces a novel Transformer-based self-supervised pre-training paradigm that pre-trains models directly on decentralized target task datasets using masked image modeling, to facilitate more robust representation learning on heterogeneous data and effective knowledge transfer to downstream models. Extensive empirical results on simulated and real-world medical imaging non-IID federated datasets show that masked image modeling with Transformers significantly improves the robustness of models against various degrees of data heterogeneity. Notably, under severe data heterogeneity, our method, without relying on any additional pre-training data, achieves an improvement of 5.06%, 1.53% and 4.58% in test accuracy on retinal, dermatology and chest X-ray classification compared to the supervised baseline with ImageNet pre-training. In addition, we show that our federated self-supervised pre-training methods yield models that generalize better to out-of-distribution data and perform more effectively when fine-tuning with limited labeled data, compared to existing FL algorithms. The code is available at https://github.com/rui-yan/SSL-FL.

Siyi Tang, Jared A. Dunnmon, Liangqiong Qu et al. · stanford

Multivariate biosignals are prevalent in many medical domains, such as electroencephalography, polysomnography, and electrocardiography. Modeling spatiotemporal dependencies in multivariate biosignals is challenging due to (1) long-range temporal dependencies and (2) complex spatial correlations between the electrodes. To address these challenges, we propose representing multivariate biosignals as time-dependent graphs and introduce GraphS4mer, a general graph neural network (GNN) architecture that improves performance on biosignal classification tasks by modeling spatiotemporal dependencies in biosignals. Specifically, (1) we leverage the Structured State Space architecture, a state-of-the-art deep sequence model, to capture long-range temporal dependencies in biosignals and (2) we propose a graph structure learning layer in GraphS4mer to learn dynamically evolving graph structures in the data. We evaluate our proposed model on three distinct biosignal classification tasks and show that GraphS4mer consistently improves over existing models, including (1) seizure detection from electroencephalographic signals, outperforming a previous GNN with self-supervised pre-training by 3.1 points in AUROC; (2) sleep staging from polysomnographic signals, a 4.1 points improvement in macro-F1 score compared to existing sleep staging models; and (3) 12-lead electrocardiogram classification, outperforming previous state-of-the-art models by 2.7 points in macro-F1 score.

Rogier van der Sluijs, Nandita Bhaskhar, Daniel Rubin et al.

Image augmentations are quintessential for effective visual representation learning across self-supervised learning techniques. While augmentation strategies for natural imaging have been studied extensively, medical images are vastly different from their natural counterparts. Thus, it is unknown whether common augmentation strategies employed in Siamese representation learning generalize to medical images and to what extent. To address this challenge, in this study, we systematically assess the effect of various augmentations on the quality and robustness of the learned representations. We train and evaluate Siamese Networks for abnormality detection on chest X-Rays across three large datasets (MIMIC-CXR, CheXpert and VinDR-CXR). We investigate the efficacy of the learned representations through experiments involving linear probing, fine-tuning, zero-shot transfer, and data efficiency. Finally, we identify a set of augmentations that yield robust representations that generalize well to both out-of-distribution data and diseases, while outperforming supervised baselines using just zero-shot transfer and linear probes by up to 20%. Our code is available at https://github.com/StanfordMIMI/siaug.

Nandita Bhaskhar, Daniel L. Rubin, Christopher Lee-Messer · stanford

Continuous monitoring of trained ML models to determine when their predictions should and should not be trusted is essential for their safe deployment. Such a framework ought to be high-performing, explainable, post-hoc and actionable. We propose TRUST-LAPSE, a "mistrust" scoring framework for continuous model monitoring. We assess the trustworthiness of each input sample's model prediction using a sequence of latent-space embeddings. Specifically, (a) our latent-space mistrust score estimates mistrust using distance metrics (Mahalanobis distance) and similarity metrics (cosine similarity) in the latent-space and (b) our sequential mistrust score determines deviations in correlations over the sequence of past input representations in a non-parametric, sliding-window based algorithm for actionable continuous monitoring. We evaluate TRUST-LAPSE via two downstream tasks: (1) distributionally shifted input detection, and (2) data drift detection. We evaluate across diverse domains - audio and vision using public datasets and further benchmark our approach on challenging, real-world electroencephalograms (EEG) datasets for seizure detection. Our latent-space mistrust scores achieve state-of-the-art results with AUROCs of 84.1 (vision), 73.9 (audio), and 77.1 (clinical EEGs), outperforming baselines by over 10 points. We expose critical failures in popular baselines that remain insensitive to input semantic content, rendering them unfit for real-world model monitoring. We show that our sequential mistrust scores achieve high drift detection rates; over 90% of the streams show < 20% error for all domains. Through extensive qualitative and quantitative evaluations, we show that our mistrust scores are more robust and provide explainability for easy adoption into practice.

Khaled Saab, Siyi Tang, Mohamed Taha et al. · stanford

A major barrier to deploying healthcare AI models is their trustworthiness. One form of trustworthiness is a model's robustness across different subgroups: while existing models may exhibit expert-level performance on aggregate metrics, they often rely on non-causal features, leading to errors in hidden subgroups. To take a step closer towards trustworthy seizure onset detection from EEG, we propose to leverage annotations that are produced by healthcare personnel in routine clinical workflows -- which we refer to as workflow notes -- that include multiple event descriptions beyond seizures. Using workflow notes, we first show that by scaling training data to an unprecedented level of 68,920 EEG hours, seizure onset detection performance significantly improves (+12.3 AUROC points) compared to relying on smaller training sets with expensive manual gold-standard labels. Second, we reveal that our binary seizure onset detection model underperforms on clinically relevant subgroups (e.g., up to a margin of 6.5 AUROC points between pediatrics and adults), while having significantly higher false positives on EEG clips showing non-epileptiform abnormalities compared to any EEG clip (+19 FPR points). To improve model robustness to hidden subgroups, we train a multilabel model that classifies 26 attributes other than seizures, such as spikes, slowing, and movement artifacts. We find that our multilabel model significantly improves overall seizure onset detection performance (+5.9 AUROC points) while greatly improving performance among subgroups (up to +8.3 AUROC points), and decreases false positives on non-epileptiform abnormalities by 8 FPR points. Finally, we propose a clinical utility metric based on false positives per 24 EEG hours and find that our multilabel model improves this clinical utility metric by a factor of 2x across different clinical settings.

Siyi Tang, Amara Tariq, Jared Dunnmon et al.

Measures to predict 30-day readmission are considered an important quality factor for hospitals as accurate predictions can reduce the overall cost of care by identifying high risk patients before they are discharged. While recent deep learning-based studies have shown promising empirical results on readmission prediction, several limitations exist that may hinder widespread clinical utility, such as (a) only patients with certain conditions are considered, (b) existing approaches do not leverage data temporality, (c) individual admissions are assumed independent of each other, which is unrealistic, (d) prior studies are usually limited to single source of data and single center data. To address these limitations, we propose a multimodal, modality-agnostic spatiotemporal graph neural network (MM-STGNN) for prediction of 30-day all-cause hospital readmission that fuses multimodal in-patient longitudinal data. By training and evaluating our methods using longitudinal chest radiographs and electronic health records from two independent centers, we demonstrate that MM-STGNN achieves AUROC of 0.79 on both primary and external datasets. Furthermore, MM-STGNN significantly outperforms the current clinical reference standard, LACE+ score (AUROC=0.61), on the primary dataset. For subset populations of patients with heart and vascular disease, our model also outperforms baselines on predicting 30-day readmission (e.g., 3.7 point improvement in AUROC in patients with heart disease). Lastly, qualitative model interpretability analysis indicates that while patients' primary diagnoses were not explicitly used to train the model, node features crucial for model prediction directly reflect patients' primary diagnoses. Importantly, our MM-STGNN is agnostic to node feature modalities and could be utilized to integrate multimodal data for triaging patients in various downstream resource allocation tasks.

Minhaj Nur Alam, Rikiya Yamashita, Vignav Ramesh et al.

Self supervised contrastive learning based pretraining allows development of robust and generalized deep learning models with small, labeled datasets, reducing the burden of label generation. This paper aims to evaluate the effect of CL based pretraining on the performance of referrable vs non referrable diabetic retinopathy (DR) classification. We have developed a CL based framework with neural style transfer (NST) augmentation to produce models with better representations and initializations for the detection of DR in color fundus images. We compare our CL pretrained model performance with two state of the art baseline models pretrained with Imagenet weights. We further investigate the model performance with reduced labeled training data (down to 10 percent) to test the robustness of the model when trained with small, labeled datasets. The model is trained and validated on the EyePACS dataset and tested independently on clinical data from the University of Illinois, Chicago (UIC). Compared to baseline models, our CL pretrained FundusNet model had higher AUC (CI) values (0.91 (0.898 to 0.930) vs 0.80 (0.783 to 0.820) and 0.83 (0.801 to 0.853) on UIC data). At 10 percent labeled training data, the FundusNet AUC was 0.81 (0.78 to 0.84) vs 0.58 (0.56 to 0.64) and 0.63 (0.60 to 0.66) in baseline models, when tested on the UIC dataset. CL based pretraining with NST significantly improves DL classification performance, helps the model generalize well (transferable from EyePACS to UIC data), and allows training with small, annotated datasets, therefore reducing ground truth annotation burden of the clinicians.

Jupinder Parmar, Khaled Saab, Brian Pogatchnik et al.

Domain generalization in medical image classification is an important problem for trustworthy machine learning to be deployed in healthcare. We find that existing approaches for domain generalization which utilize ground-truth abnormality segmentations to control feature attributions have poor out-of-distribution (OOD) performance relative to the standard baseline of empirical risk minimization (ERM). We investigate what regions of an image are important for medical image classification and show that parts of the background, that which is not contained in the abnormality segmentation, provides helpful signal. We then develop a new task-specific mask which covers all relevant regions. Utilizing this new segmentation mask significantly improves the performance of the existing methods on the OOD test sets. To obtain better generalization results than ERM, we find it necessary to scale up the training data size in addition to the usage of these task-specific masks.

Yan-Ran, Wang, Liangqiong Qu et al.

Despite its tremendous value for the diagnosis, treatment monitoring and surveillance of children with cancer, whole body staging with positron emission tomography (PET) is time consuming and associated with considerable radiation exposure. 100x (1% of the standard clinical dosage) ultra-low-dose/ultra-fast whole-body PET reconstruction has the potential for cancer imaging with unprecedented speed and improved safety, but it cannot be achieved by the naive use of machine learning techniques. In this study, we utilize the global similarity between baseline and follow-up PET and magnetic resonance (MR) images to develop Masked-LMCTrans, a longitudinal multi-modality co-attentional CNN-Transformer that provides interaction and joint reasoning between serial PET/MRs of the same patient. We mask the tumor area in the referenced baseline PET and reconstruct the follow-up PET scans. In this manner, Masked-LMCTrans reconstructs 100x almost-zero radio-exposure whole-body PET that was not possible before. The technique also opens a new pathway for longitudinal radiology imaging reconstruction, a significantly under-explored area to date. Our model was trained and tested with Stanford PET/MRI scans of pediatric lymphoma patients and evaluated externally on PET/MRI images from Tübingen University. The high image quality of the reconstructed 100x whole-body PET images resulting from the application of Masked-LMCTrans will substantially advance the development of safer imaging approaches and shorter exam-durations for pediatric patients, as well as expand the possibilities for frequent longitudinal monitoring of these patients by PET.

Liangqiong Qu, Yuyin Zhou, Paul Pu Liang et al.

Federated learning is an emerging research paradigm enabling collaborative training of machine learning models among different organizations while keeping data private at each institution. Despite recent progress, there remain fundamental challenges such as the lack of convergence and the potential for catastrophic forgetting across real-world heterogeneous devices. In this paper, we demonstrate that self-attention-based architectures (e.g., Transformers) are more robust to distribution shifts and hence improve federated learning over heterogeneous data. Concretely, we conduct the first rigorous empirical investigation of different neural architectures across a range of federated algorithms, real-world benchmarks, and heterogeneous data splits. Our experiments show that simply replacing convolutional networks with Transformers can greatly reduce catastrophic forgetting of previous devices, accelerate convergence, and reach a better global model, especially when dealing with heterogeneous data. We release our code and pretrained models at https://github.com/Liangqiong/ViT-FL-main to encourage future exploration in robust architectures as an alternative to current research efforts on the optimization front.

Vignav Ramesh, Blaine Rister, Daniel L. Rubin

Chest X-rays of coronavirus disease 2019 (COVID-19) patients are frequently obtained to determine the extent of lung disease and are a valuable source of data for creating artificial intelligence models. Most work to date assessing disease severity on chest imaging has focused on segmenting computed tomography (CT) images; however, given that CTs are performed much less frequently than chest X-rays for COVID-19 patients, automated lung lesion segmentation on chest X-rays could be clinically valuable. There currently exists a universal shortage of chest X-rays with ground truth COVID-19 lung lesion annotations, and manually contouring lung opacities is a tedious, labor-intensive task. To accelerate severity detection and augment the amount of publicly available chest X-ray training data for supervised deep learning (DL) models, we leverage existing annotated CT images to generate frontal projection "chest X-ray" images for training COVID-19 chest X-ray models. In this paper, we propose an automated pipeline for segmentation of COVID-19 lung lesions on chest X-rays comprised of a Mask R-CNN trained on a mixed dataset of open-source chest X-rays and coronal X-ray projections computed from annotated volumetric CTs. On a test set containing 40 chest X-rays of COVID-19 positive patients, our model achieved IoU scores of 0.81 $\pm$ 0.03 and 0.79 $\pm$ 0.03 when trained on a dataset of 60 chest X-rays and on a mixed dataset of 10 chest X-rays and 50 projections from CTs, respectively. Our model far outperforms current baselines with limited supervised training and may assist in automated COVID-19 severity quantification on chest X-rays.

Ali Hatamizadeh, Assaf Hoogi, Debleena Sengupta et al.

Lesion segmentation is an important problem in computer-assisted diagnosis that remains challenging due to the prevalence of low contrast, irregular boundaries that are unamenable to shape priors. We introduce Deep Active Lesion Segmentation (DALS), a fully automated segmentation framework for that leverages the powerful nonlinear feature extraction abilities of fully Convolutional Neural Networks (CNNs) and the precise boundary delineation abilities of Active Contour Models (ACMs). Our DALS framework benefits from an improved level-set ACM formulation with a per-pixel-parameterized energy functional and a novel multiscale encoder-decoder CNN that learns an initialization probability map along with parameter maps for the ACM. We evaluate our lesion segmentation model on a new Multiorgan Lesion Segmentation (MLS) dataset that contains images of various organs, including brain, liver, and lung, across different imaging modalities---MR and CT. Our results demonstrate favorable performance compared to competing methods, especially for small training datasets. Source code : $\text{https://github.com/ahatamiz/dals}$

Balasubramanian Narasimhan, Daniel L. Rubin, Samuel M. Gross et al.

Bringing together the information latent in distributed medical databases promises to personalize medical care by enabling reliable, stable modeling of outcomes with rich feature sets (including patient characteristics and treatments received). However, there are barriers to aggregation of medical data, due to lack of standardization of ontologies, privacy concerns, proprietary attitudes toward data, and a reluctance to give up control over end use. Aggregation of data is not always necessary for model fitting. In models based on maximizing a likelihood, the computations can be distributed, with aggregation limited to the intermediate results of calculations on local data, rather than raw data. Distributed fitting is also possible for singular value decomposition. There has been work on the technical aspects of shared computation for particular applications, but little has been published on the software needed to support the "social networking" aspect of shared computing, to reduce the barriers to collaboration. We describe a set of software tools that allow the rapid assembly of a collaborative computational project, based on the flexible and extensible R statistical software and other open source packages, that can work across a heterogeneous collection of database environments, with full transparency to allow local officials concerned with privacy protections to validate the safety of the method. We describe the principles, architecture, and successful test results for the site-stratified Cox model and rank-k Singular Value Decomposition (SVD).

Alexander Berdichevsky, Mor Peleg, Daniel L. Rubin

Objective. Mammography reports document the diagnosis of patients' conditions. However, many reports contain non-standard terms (non-BI-RADS descriptors) and incomplete statements, which can lead to conclusions that are not well-supported by the reported findings. Our aim was to develop a tool to detect such discrepancies by comparing the reported conclusions to those that would be expected based on the reported radiology findings. Materials and Methods. A deidentified data set from an academic hospital containing 258 mammography reports supplemented by 120 reports found on the web was used for training and evaluation. Spell checking and term normalization was used to unambiguously determine the reported BI-RADS descriptors. The resulting data were input into seven classifiers that classify mammography reports, based on their Findings sections, into seven BI-RADS final assessment categories. Finally, the semantic similarity score of a report to each BI-RADS category is reported. Results. Our term normalization algorithm correctly identified 97% of the BI-RADS descriptors in mammography reports. Our system provided 76% precision and 83% recall in correctly classifying the reports according to BI-RADS final assessment category. Discussion. The strength of our approach relies on providing high importance to BI-RADS terms in the summarization phase, on the semantic similarity that considers the complex data representation, and on the classification into all seven BI-RADs categories. Conclusion. BI-RADS descriptors and expected final assessment categories could be automatically detected by our approach with fairly good accuracy, which could be used to make users aware that their reported findings do not match well with their conclusion.

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

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

Yuyin Zhou, Shih-Cheng Huang, Jason Alan Fries et al.

Despite the routine use of electronic health record (EHR) data by radiologists to contextualize clinical history and inform image interpretation, the majority of deep learning architectures for medical imaging are unimodal, i.e., they only learn features from pixel-level information. Recent research revealing how race can be recovered from pixel data alone highlights the potential for serious biases in models which fail to account for demographics and other key patient attributes. Yet the lack of imaging datasets which capture clinical context, inclusive of demographics and longitudinal medical history, has left multimodal medical imaging underexplored. To better assess these challenges, we present RadFusion, a multimodal, benchmark dataset of 1794 patients with corresponding EHR data and high-resolution computed tomography (CT) scans labeled for pulmonary embolism. We evaluate several representative multimodal fusion models and benchmark their fairness properties across protected subgroups, e.g., gender, race/ethnicity, age. Our results suggest that integrating imaging and EHR data can improve classification performance and robustness without introducing large disparities in the true positive rate between population groups.

Xiang Bai, Hanchen Wang, Liya Ma et al.

Artificial intelligence (AI) provides a promising substitution for streamlining COVID-19 diagnoses. However, concerns surrounding security and trustworthiness impede the collection of large-scale representative medical data, posing a considerable challenge for training a well-generalised model in clinical practices. To address this, we launch the Unified CT-COVID AI Diagnostic Initiative (UCADI), where the AI model can be distributedly trained and independently executed at each host institution under a federated learning framework (FL) without data sharing. Here we show that our FL model outperformed all the local models by a large yield (test sensitivity /specificity in China: 0.973/0.951, in the UK: 0.730/0.942), achieving comparable performance with a panel of professional radiologists. We further evaluated the model on the hold-out (collected from another two hospitals leaving out the FL) and heterogeneous (acquired with contrast materials) data, provided visual explanations for decisions made by the model, and analysed the trade-offs between the model performance and the communication costs in the federated training process. Our study is based on 9,573 chest computed tomography scans (CTs) from 3,336 patients collected from 23 hospitals located in China and the UK. Collectively, our work advanced the prospects of utilising federated learning for privacy-preserving AI in digital health.

Liangqiong Qu, Niranjan Balachandar, Daniel L Rubin

Federated learning enables multiple institutions to collaboratively train machine learning models on their local data in a privacy-preserving way. However, its distributed nature often leads to significant heterogeneity in data distributions across institutions. In this paper, we investigate the deleterious impact of a taxonomy of data heterogeneity regimes on federated learning methods, including quantity skew, label distribution skew, and imaging acquisition skew. We show that the performance degrades with the increasing degrees of data heterogeneity. We present several mitigation strategies to overcome performance drops from data heterogeneity, including weighted average for data quantity skew, weighted loss and batch normalization averaging for label distribution skew. The proposed optimizations to federated learning methods improve their capability of handling heterogeneity across institutions, which provides valuable guidance for the deployment of federated learning in real clinical applications.

Miao Zhang, Liangqiong Qu, Praveer Singh et al.

Federated learning is an emerging research paradigm for enabling collaboratively training deep learning models without sharing patient data. However, the data from different institutions are usually heterogeneous across institutions, which may reduce the performance of models trained using federated learning. In this study, we propose a novel heterogeneity-aware federated learning method, SplitAVG, to overcome the performance drops from data heterogeneity in federated learning. Unlike previous federated methods that require complex heuristic training or hyper parameter tuning, our SplitAVG leverages the simple network split and feature map concatenation strategies to encourage the federated model training an unbiased estimator of the target data distribution. We compare SplitAVG with seven state-of-the-art federated learning methods, using centrally hosted training data as the baseline on a suite of both synthetic and real-world federated datasets. We find that the performance of models trained using all the comparison federated learning methods degraded significantly with the increasing degrees of data heterogeneity. In contrast, SplitAVG method achieves comparable results to the baseline method under all heterogeneous settings, that it achieves 96.2% of the accuracy and 110.4% of the mean absolute error obtained by the baseline in a diabetic retinopathy binary classification dataset and a bone age prediction dataset, respectively, on highly heterogeneous data partitions. We conclude that SplitAVG method can effectively overcome the performance drops from variability in data distributions across institutions. Experimental results also show that SplitAVG can be adapted to different base networks and generalized to various types of medical imaging tasks.

Liangqiong Qu, Niranjan Balachandar, Miao Zhang et al.

Collaborative learning, which enables collaborative and decentralized training of deep neural networks at multiple institutions in a privacy-preserving manner, is rapidly emerging as a valuable technique in healthcare applications. However, its distributed nature often leads to significant heterogeneity in data distributions across institutions. In this paper, we present a novel generative replay strategy to address the challenge of data heterogeneity in collaborative learning methods. Different from traditional methods that directly aggregating the model parameters, we leverage generative adversarial learning to aggregate the knowledge from all the local institutions. Specifically, instead of directly training a model for task performance, we develop a novel dual model architecture: a primary model learns the desired task, and an auxiliary "generative replay model" allows aggregating knowledge from the heterogenous clients. The auxiliary model is then broadcasted to the central sever, to regulate the training of primary model with an unbiased target distribution. Experimental results demonstrate the capability of the proposed method in handling heterogeneous data across institutions. On highly heterogeneous data partitions, our model achieves ~4.88% improvement in the prediction accuracy on a diabetic retinopathy classification dataset, and ~49.8% reduction of mean absolution value on a Bone Age prediction dataset, respectively, compared to the state-of-the art collaborative learning methods.

Florian Dubost, Erin Hong, Max Pike et al.

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

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

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

Sharut Gupta, Praveer Singh, Ken Chang et al.

Model brittleness is a key concern when deploying deep learning models in real-world medical settings. A model that has high performance at one institution may suffer a significant decline in performance when tested at other institutions. While pooling datasets from multiple institutions and retraining may provide a straightforward solution, it is often infeasible and may compromise patient privacy. An alternative approach is to fine-tune the model on subsequent institutions after training on the original institution. Notably, this approach degrades model performance at the original institution, a phenomenon known as catastrophic forgetting. In this paper, we develop an approach to address catastrophic forget-ting based on elastic weight consolidation combined with modulation of batch normalization statistics under two scenarios: first, for expanding the domain from one imaging system's data to another imaging system's, and second, for expanding the domain from a large multi-institutional dataset to another single institution dataset. We show that our approach outperforms several other state-of-the-art approaches and provide theoretical justification for the efficacy of batch normalization modulation. The results of this study are generally applicable to the deployment of any clinical deep learning model which requires domain expansion.

Rikiya Yamashita, Jin Long, Snikitha Banda et al.

Suboptimal generalization of machine learning models on unseen data is a key challenge which hampers the clinical applicability of such models to medical imaging. Although various methods such as domain adaptation and domain generalization have evolved to combat this challenge, learning robust and generalizable representations is core to medical image understanding, and continues to be a problem. Here, we propose STRAP (Style TRansfer Augmentation for histoPathology), a form of data augmentation based on random style transfer from non-medical style source such as artistic paintings, for learning domain-agnostic visual representations in computational pathology. Style transfer replaces the low-level texture content of an image with the uninformative style of randomly selected style source image, while preserving the original high-level semantic content. This improves robustness to domain shift and can be used as a simple yet powerful tool for learning domain-agnostic representations. We demonstrate that STRAP leads to state-of-the-art performance, particularly in the presence of domain shifts, on two particular classification tasks in computational pathology.

Florian Dubost, Erin Hong, Nandita Bhaskhar et al.

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

Sharut Gupta, Praveer Singh, Ken Chang et al.

Model brittleness is a primary concern when deploying deep learning models in medical settings owing to inter-institution variations, like patient demographics and intra-institution variation, such as multiple scanner types. While simply training on the combined datasets is fraught with data privacy limitations, fine-tuning the model on subsequent institutions after training it on the original institution results in a decrease in performance on the original dataset, a phenomenon called catastrophic forgetting. In this paper, we investigate trade-off between model refinement and retention of previously learned knowledge and subsequently address catastrophic forgetting for the assessment of mammographic breast density. More specifically, we propose a simple yet effective approach, adapting Elastic weight consolidation (EWC) using the global batch normalization (BN) statistics of the original dataset. The results of this study provide guidance for the deployment of clinical deep learning models where continuous learning is needed for domain expansion.

Blaine Rister, Daniel L. Rubin

Neuron death is a complex phenomenon with implications for model trainability: the deeper the network, the lower the probability of finding a valid initialization. In this work, we derive both upper and lower bounds on the probability that a ReLU network is initialized to a trainable point, as a function of model hyperparameters. We show that it is possible to increase the depth of a network indefinitely, so long as the width increases as well. Furthermore, our bounds are asymptotically tight under reasonable assumptions: first, the upper bound coincides with the true probability for a single-layer network with the largest possible input set. Second, the true probability converges to our lower bound as the input set shrinks to a single point, or as the network complexity grows under an assumption about the output variance. We confirm these results by numerical simulation, showing rapid convergence to the lower bound with increasing network depth. Then, motivated by the theory, we propose a practical sign flipping scheme which guarantees that the ratio of living data points in a $k$-layer network is at least $2^{-k}$. Finally, we show how these issues are mitigated by network design features currently seen in practice, such as batch normalization, residual connections, dense networks and skip connections. This suggests that neuron death may provide insight into the efficacy of various model architectures.

Darvin Yi, Endre Grøvik, Michael Iv et al.

We introduce a novel ensembling method, Random Bundle (RB), that improves performance for brain metastases segmentation. We create our ensemble by training each network on our dataset with 50% of our annotated lesions censored out. We also apply a lopsided bootstrap loss to recover performance after inducing an in silico 50% false negative rate and make our networks more sensitive. We improve our network detection of lesions's mAP value by 39% and more than triple the sensitivity at 80% precision. We also show slight improvements in segmentation quality through DICE score. Further, RB ensembling improves performance over baseline by a larger margin than a variety of popular ensembling strategies. Finally, we show that RB ensembling is computationally efficient by comparing its performance to a single network when both systems are constrained to have the same compute.

Darvin Yi, Endre Grøvik, Michael Iv et al.

Deep learning has proven to be an essential tool for medical image analysis. However, the need for accurately labeled input data, often requiring time- and labor-intensive annotation by experts, is a major limitation to the use of deep learning. One solution to this challenge is to allow for use of coarse or noisy labels, which could permit more efficient and scalable labeling of images. In this work, we develop a lopsided loss function based on entropy regularization that assumes the existence of a nontrivial false negative rate in the target annotations. Starting with a carefully annotated brain metastasis lesion dataset, we simulate data with false negatives by (1) randomly censoring the annotated lesions and (2) systematically censoring the smallest lesions. The latter better models true physician error because smaller lesions are harder to notice than the larger ones. Even with a simulated false negative rate as high as 50%, applying our loss function to randomly censored data preserves maximum sensitivity at 97% of the baseline with uncensored training data, compared to just 10% for a standard loss function. For the size-based censorship, performance is restored from 17% with the current standard to 88% with our lopsided bootstrap loss. Our work will enable more efficient scaling of the image labeling process, in parallel with other approaches on creating more efficient user interfaces and tools for annotation.

Endre Grøvik, Darvin Yi, Michael Iv et al.

The purpose was to assess the clinical value of a novel DropOut model for detecting and segmenting brain metastases, in which a neural network is trained on four distinct MRI sequences using an input dropout layer, thus simulating the scenario of missing MRI data by training on the full set and all possible subsets of the input data. This retrospective, multi-center study, evaluated 165 patients with brain metastases. A deep learning based segmentation model for automatic segmentation of brain metastases, named DropOut, was trained on multi-sequence MRI from 100 patients, and validated/tested on 10/55 patients. The segmentation results were compared with the performance of a state-of-the-art DeepLabV3 model. The MR sequences in the training set included pre- and post-gadolinium (Gd) T1-weighted 3D fast spin echo, post-Gd T1-weighted inversion recovery (IR) prepped fast spoiled gradient echo, and 3D fluid attenuated inversion recovery (FLAIR), whereas the test set did not include the IR prepped image-series. The ground truth were established by experienced neuroradiologists. The results were evaluated using precision, recall, Dice score, and receiver operating characteristics (ROC) curve statistics, while the Wilcoxon rank sum test was used to compare the performance of the two neural networks. The area under the ROC curve (AUC), averaged across all test cases, was 0.989+-0.029 for the DropOut model and 0.989+-0.023 for the DeepLabV3 model (p=0.62). The DropOut model showed a significantly higher Dice score compared to the DeepLabV3 model (0.795+-0.105 vs. 0.774+-0.104, p=0.017), and a significantly lower average false positive rate of 3.6/patient vs. 7.0/patient (p<0.001) using a 10mm3 lesion-size limit. The DropOut model may facilitate accurate detection and segmentation of brain metastases on a multi-center basis, even when the test cohort is missing MRI input data.

Darvin Yi, Endre Grøvik, Michael Iv et al.

Magnetic resonance (MR) imaging is an essential diagnostic tool in clinical medicine. Recently, a variety of deep learning methods have been applied to segmentation tasks in medical images, with promising results for computer-aided diagnosis. For MR images, effectively integrating different pulse sequences is important to optimize performance. However, the best way to integrate different pulse sequences remains unclear. In this study, we evaluate multiple architectural features and characterize their effects in the task of metastasis segmentation. Specifically, we consider (1) different pulse sequence integration schemas, (2) different modes of weight sharing for parallel network branches, and (3) a new approach for enabling robustness to missing pulse sequences. We find that levels of integration and modes of weight sharing that favor low variance work best in our regime of small data (n = 100). By adding an input-level dropout layer, we could preserve the overall performance of these networks while allowing for inference on inputs with missing pulse sequence. We illustrate not only the generalizability of the network but also the utility of this robustness when applying the trained model to data from a different center, which does not use the same pulse sequences. Finally, we apply network visualization methods to better understand which input features are most important for network performance. Together, these results provide a framework for building networks with enhanced robustness to missing data while maintaining comparable performance in medical imaging applications.

Okyaz Eminaga, Yuri Tolkach, Christian Kunder et al.

The current study detects different morphologies related to prostate pathology using deep learning models; these models were evaluated on 2,121 hematoxylin and eosin (H&E) stain histology images captured using bright field microscopy, which spanned a variety of image qualities, origins (whole slide, tissue micro array, whole mount, Internet), scanning machines, timestamps, H&E staining protocols, and institutions. For case usage, these models were applied for the annotation tasks in clinician-oriented pathology reports for prostatectomy specimens. The true positive rate (TPR) for slides with prostate cancer was 99.7% by a false positive rate of 0.785%. The F1-scores of Gleason patterns reported in pathology reports ranged from 0.795 to 1.0 at the case level. TPR was 93.6% for the cribriform morphology and 72.6% for the ductal morphology. The correlation between the ground truth and the prediction for the relative tumor volume was 0.987 n. Our models cover the major components of prostate pathology and successfully accomplish the annotation tasks.

Okyaz Eminaga, Mahmoud Abbas, Christian Kunder et al.

Different convolutional neural network (CNN) models have been tested for their application in histological image analyses. However, these models are prone to overfitting due to their large parameter capacity, requiring more data or valuable computational resources for model training. Given these limitations, we introduced a novel architecture (termed PlexusNet). We utilized 310 Hematoxylin and Eosin stained (H&E) annotated histological images of prostate cancer cases from TCGA-PRAD and Stanford University and 398 H&E whole slides images from the Camelyon 2016 challenge. PlexusNet-architecture -derived models were compared to models derived from several existing "state of the art" architectures. We measured discrimination accuracy, calibration, and clinical utility. An ablation study was conducted to study the effect of each component of PlexusNet on model performance. A well-fitted PlexusNet-based model delivered comparable classification performance (AUC: 0.963) in distinguishing prostate cancer from healthy tissues, although it was at least 23 times smaller, had a better model calibration and clinical utility than the comparison models. A separate smaller PlexusNet model accurately detected slides with breast cancer metastases (AUC: 0.978); it helped reduce the slide number to examine by 43.8% without consequences, although its parameter capacity was 200 times smaller than ResNet18. We found that the partitioning of the development set influences the model calibration for all models. However, with PlexusNet architecture, we could achieve comparable well-calibrated models trained on different partitions. In conclusion, PlexusNet represents a novel model architecture for histological image analysis that achieves classification performance comparable to other models while providing orders-of-magnitude parameter reduction.

Assaf Hoogi, Brian Wilcox, Yachee Gupta et al.

We propose a novel architecture for object classification, called Self-Attention Capsule Networks (SACN). SACN is the first model that incorporates the Self-Attention mechanism as an integral layer within the Capsule Network (CapsNet). While the Self-Attention mechanism supplies a long-range dependencies, results in selecting the more dominant image regions to focus on, the CapsNet analyzes the relevant features and their spatial correlations inside these regions only. The features are extracted in the convolutional layer. Then, the Self-Attention layer learns to suppress irrelevant regions based on features analysis and highlights salient features useful for a specific task. The attention map is then fed into the CapsNet primary layer that is followed by a classification layer. The proposed SACN model was designed to solve two main limitations of the baseline CapsNet - analysis of complex data and significant computational load. In this work, we use a shallow CapsNet architecture and compensates for the absence of a deeper network by using the Self-Attention module to significantly improve the results. The proposed Self-Attention CapsNet architecture was extensively evaluated on six different datasets, mainly on three different medical sets, in addition to the natural MNIST, SVHN and CIFAR10. The model was able to classify images and their patches with diverse and complex backgrounds better than the baseline CapsNet. As a result, the proposed Self-Attention CapsNet significantly improved classification performance within and across different datasets and outperformed the baseline CapsNet, ResNet-18 and DenseNet-40 not only in classification accuracy but also in robustness.

Jared Dunnmon, Alexander Ratner, Nishith Khandwala et al.

Labeling training datasets has become a key barrier to building medical machine learning models. One strategy is to generate training labels programmatically, for example by applying natural language processing pipelines to text reports associated with imaging studies. We propose cross-modal data programming, which generalizes this intuitive strategy in a theoretically-grounded way that enables simpler, clinician-driven input, reduces required labeling time, and improves with additional unlabeled data. In this approach, clinicians generate training labels for models defined over a target modality (e.g. images or time series) by writing rules over an auxiliary modality (e.g. text reports). The resulting technical challenge consists of estimating the accuracies and correlations of these rules; we extend a recent unsupervised generative modeling technique to handle this cross-modal setting in a provably consistent way. Across four applications in radiography, computed tomography, and electroencephalography, and using only several hours of clinician time, our approach matches or exceeds the efficacy of physician-months of hand-labeling with statistical significance, demonstrating a fundamentally faster and more flexible way of building machine learning models in medicine.

Imon Banerjee, Luis de Sisternes, Joelle Hallak et al.

We propose a hybrid sequential deep learning model to predict the risk of AMD progression in non-exudative AMD eyes at multiple timepoints, starting from short-term progression (3-months) up to long-term progression (21-months). Proposed model combines radiomics and deep learning to handle challenges related to imperfect ratio of OCT scan dimension and training cohort size. We considered a retrospective clinical trial dataset that includes 671 fellow eyes with 13,954 dry AMD observations for training and validating the machine learning models on a 10-fold cross validation setting. The proposed RNN model achieved high accuracy (0.96 AUCROC) for the prediction of both short term and long-term AMD progression, and outperformed the traditional random forest model trained. High accuracy achieved by the RNN establishes the ability to identify AMD patients at risk of progressing to advanced AMD at an early stage which could have a high clinical impact as it allows for optimal clinical follow-up, with more frequent screening and potential earlier treatment for those patients at high risk.

Blaine Rister, Darvin Yi, Kaushik Shivakumar et al.

Fully-convolutional neural networks have achieved superior performance in a variety of image segmentation tasks. However, their training requires laborious manual annotation of large datasets, as well as acceleration by parallel processors with high-bandwidth memory, such as GPUs. We show that simple models can achieve competitive accuracy for organ segmentation on CT images when trained with extensive data augmentation, which leverages existing graphics hardware to quickly apply geometric and photometric transformations to 3D image data. On 3 mm^3 CT volumes, our GPU implementation is 2.6-8X faster than a widely-used CPU version, including communication overhead. We also show how to automatically generate training labels using rudimentary morphological operations, which are efficiently computed by 3D Fourier transforms. We combined fully-automatic labels for the lungs and bone with semi-automatic ones for the liver, kidneys and bladder, to create a dataset of 130 labeled CT scans. To achieve the best results from data augmentation, our model uses the intersection-over-union (IOU) loss function, a close relative of the Dice loss. We discuss its mathematical properties and explain why it outperforms the usual weighted cross-entropy loss for unbalanced segmentation tasks. We conclude that there is no unique IOU loss function, as the naive one belongs to a broad family of functions with the same essential properties. When combining data augmentation with the IOU loss, our model achieves a Dice score of 78-92% for each organ. The trained model, code and dataset will be made publicly available, to further medical imaging research.

Haque Ishfaq, Assaf Hoogi, Daniel Rubin

Deep metric learning has been demonstrated to be highly effective in learning semantic representation and encoding information that can be used to measure data similarity, by relying on the embedding learned from metric learning. At the same time, variational autoencoder (VAE) has widely been used to approximate inference and proved to have a good performance for directed probabilistic models. However, for traditional VAE, the data label or feature information are intractable. Similarly, traditional representation learning approaches fail to represent many salient aspects of the data. In this project, we propose a novel integrated framework to learn latent embedding in VAE by incorporating deep metric learning. The features are learned by optimizing a triplet loss on the mean vectors of VAE in conjunction with standard evidence lower bound (ELBO) of VAE. This approach, which we call Triplet based Variational Autoencoder (TVAE), allows us to capture more fine-grained information in the latent embedding. Our model is tested on MNIST data set and achieves a high triplet accuracy of 95.60% while the traditional VAE (Kingma & Welling, 2013) achieves triplet accuracy of 75.08%.

Imon Banerjee, Michael Francis Gensheimer, Douglas J. Wood et al.

We propose a deep learning model - Probabilistic Prognostic Estimates of Survival in Metastatic Cancer Patients (PPES-Met) for estimating short-term life expectancy (3 months) of the patients by analyzing free-text clinical notes in the electronic medical record, while maintaining the temporal visit sequence. In a single framework, we integrated semantic data mapping and neural embedding technique to produce a text processing method that extracts relevant information from heterogeneous types of clinical notes in an unsupervised manner, and we designed a recurrent neural network to model the temporal dependency of the patient visits. The model was trained on a large dataset (10,293 patients) and validated on a separated dataset (1818 patients). Our method achieved an area under the ROC curve (AUC) of 0.89. To provide explain-ability, we developed an interactive graphical tool that may improve physician understanding of the basis for the model's predictions. The high accuracy and explain-ability of the PPES-Met model may enable our model to be used as a decision support tool to personalize metastatic cancer treatment and provide valuable assistance to the physicians.

Imon Banerjee, Sriraman Madhavan, Roger Eric Goldman et al.

Radiology reports are a rich resource for advancing deep learning applications in medicine by leveraging the large volume of data continuously being updated, integrated, and shared. However, there are significant challenges as well, largely due to the ambiguity and subtlety of natural language. We propose a hybrid strategy that combines semantic-dictionary mapping and word2vec modeling for creating dense vector embeddings of free-text radiology reports. Our method leverages the benefits of both semantic-dictionary mapping as well as unsupervised learning. Using the vector representation, we automatically classify the radiology reports into three classes denoting confidence in the diagnosis of intracranial hemorrhage by the interpreting radiologist. We performed experiments with varying hyperparameter settings of the word embeddings and a range of different classifiers. Best performance achieved was a weighted precision of 88% and weighted recall of 90%. Our work offers the potential to leverage unstructured electronic health record data by allowing direct analysis of narrative clinical notes.

Ken Chang, Niranjan Balachandar, Carson K Lam et al.

Deep learning has become a promising approach for automated medical diagnoses. When medical data samples are limited, collaboration among multiple institutions is necessary to achieve high algorithm performance. However, sharing patient data often has limitations due to technical, legal, or ethical concerns. In such cases, sharing a deep learning model is a more attractive alternative. The best method of performing such a task is unclear, however. In this study, we simulate the dissemination of learning deep learning network models across four institutions using various heuristics and compare the results with a deep learning model trained on centrally hosted patient data. The heuristics investigated include ensembling single institution models, single weight transfer, and cyclical weight transfer. We evaluated these approaches for image classification in three independent image collections (retinal fundus photos, mammography, and ImageNet). We find that cyclical weight transfer resulted in a performance (testing accuracy = 77.3%) that was closest to that of centrally hosted patient data (testing accuracy = 78.7%). We also found that there is an improvement in the performance of cyclical weight transfer heuristic with high frequency of weight transfer.

Paroma Varma, Bryan He, Payal Bajaj et al.

Obtaining enough labeled data to robustly train complex discriminative models is a major bottleneck in the machine learning pipeline. A popular solution is combining multiple sources of weak supervision using generative models. The structure of these models affects training label quality, but is difficult to learn without any ground truth labels. We instead rely on these weak supervision sources having some structure by virtue of being encoded programmatically. We present Coral, a paradigm that infers generative model structure by statically analyzing the code for these heuristics, thus reducing the data required to learn structure significantly. We prove that Coral's sample complexity scales quasilinearly with the number of heuristics and number of relations found, improving over the standard sample complexity, which is exponential in $n$ for identifying $n^{\textrm{th}}$ degree relations. Experimentally, Coral matches or outperforms traditional structure learning approaches by up to 3.81 F1 points. Using Coral to model dependencies instead of assuming independence results in better performance than a fully supervised model by 3.07 accuracy points when heuristics are used to label radiology data without ground truth labels.

Darvin Yi, Rebecca Lynn Sawyer, David Cohn et al.

Breast cancer has the highest incidence and second highest mortality rate for women in the US. Our study aims to utilize deep learning for benign/malignant classification of mammogram tumors using a subset of cases from the Digital Database of Screening Mammography (DDSM). Though it was a small dataset from the view of Deep Learning (about 1000 patients), we show that currently state of the art architectures of deep learning can find a robust signal, even when trained from scratch. Using convolutional neural networks (CNNs), we are able to achieve an accuracy of 85% and an ROC AUC of 0.91, while leading hand-crafted feature based methods are only able to achieve an accuracy of 71%. We investigate an amalgamation of architectures to show that our best result is reached with an ensemble of the lightweight GoogLe Nets tasked with interpreting both the coronal caudal view and the mediolateral oblique view, simply averaging the probability scores of both views to make the final prediction. In addition, we have created a novel method to visualize what features the neural network detects for the benign/malignant classification, and have correlated those features with well known radiological features, such as spiculation. Our algorithm significantly improves existing classification methods for mammography lesions and identifies features that correlate with established clinical markers.

Assaf Hoogi, John W. Lambert, Yefeng Zheng et al.

We propose a fully-automated method for accurate and robust detection and segmentation of potentially cancerous lesions found in the liver and in lymph nodes. The process is performed in three steps, including organ detection, lesion detection and lesion segmentation. Our method applies machine learning techniques such as marginal space learning and convolutional neural networks, as well as active contour models. The method proves to be robust in its handling of extremely high lesion diversity. We tested our method on volumetric computed tomography (CT) images, including 42 volumes containing liver lesions and 86 volumes containing 595 pathological lymph nodes. Preliminary results under 10-fold cross validation show that for both the liver lesions and the lymph nodes, a total detection sensitivity of 0.53 and average Dice score of $0.71 \pm 0.15$ for segmentation were obtained.

Imon Banerjee, Lewis Hahn, Geoffrey Sonn et al.

We propose an automated method for detecting aggressive prostate cancer(CaP) (Gleason score >=7) based on a comprehensive analysis of the lesion and the surrounding normal prostate tissue which has been simultaneously captured in T2-weighted MR images, diffusion-weighted images (DWI) and apparent diffusion coefficient maps (ADC). The proposed methodology was tested on a dataset of 79 patients (40 aggressive, 39 non-aggressive). We evaluated the performance of a wide range of popular quantitative imaging features on the characterization of aggressive versus non-aggressive CaP. We found that a group of 44 discriminative predictors among 1464 quantitative imaging features can be used to produce an area under the ROC curve of 0.73.

Blaine Rister, Daniel L Rubin

Although artificial neural networks have shown great promise in applications including computer vision and speech recognition, there remains considerable practical and theoretical difficulty in optimizing their parameters. The seemingly unreasonable success of gradient descent methods in minimizing these non-convex functions remains poorly understood. In this work we offer some theoretical guarantees for networks with piecewise affine activation functions, which have in recent years become the norm. We prove three main results. Firstly, that the network is piecewise convex as a function of the input data. Secondly, that the network, considered as a function of the parameters in a single layer, all others held constant, is again piecewise convex. Finally, that the network as a function of all its parameters is piecewise multi-convex, a generalization of biconvexity. From here we characterize the local minima and stationary points of the training objective, showing that they minimize certain subsets of the parameter space. We then analyze the performance of two optimization algorithms on multi-convex problems: gradient descent, and a method which repeatedly solves a number of convex sub-problems. We prove necessary convergence conditions for the first algorithm and both necessary and sufficient conditions for the second, after introducing regularization to the objective. Finally, we remark on the remaining difficulty of the global optimization problem. Under the squared error objective, we show that by varying the training data, a single rectifier neuron admits local minima arbitrarily far apart, both in objective value and parameter space.

Assaf Hoogi, Christopher F. Beaulieu, Guilherme M. Cunha et al.

We propose a novel method, the adaptive local window, for improving level set segmentation technique. The window is estimated separately for each contour point, over iterations of the segmentation process, and for each individual object. Our method considers the object scale, the spatial texture, and changes of the energy functional over iterations. Global and local statistics are considered by calculating several gray level co-occurrence matrices. We demonstrate the capabilities of the method in the domain of medical imaging for segmenting 233 images with liver lesions. To illustrate the strength of our method, those images were obtained by either Computed Tomography or Magnetic Resonance Imaging. Moreover, we analyzed images using three different energy models. We compare our method to a global level set segmentation and to local framework that uses predefined fixed-size square windows. The results indicate that our proposed method outperforms the other methods in terms of agreement with the manual marking and dependence on contour initialization or the energy model used. In case of complex lesions, such as low contrast lesions, heterogeneous lesions, or lesions with a noisy background, our method shows significantly better segmentation with an improvement of 0.25+- 0.13 in Dice similarity coefficient, compared with state of the art fixed-size local windows (Wilcoxon, p < 0.001).

Aaron Adcock, Daniel Rubin, Gunnar Carlsson

In this paper we present a methodology of classifying hepatic (liver) lesions using multidimensional persistent homology, the matching metric (also called the bottleneck distance), and a support vector machine. We present our classification results on a dataset of 132 lesions that have been outlined and annotated by radiologists. We find that topological features are useful in the classification of hepatic lesions. We also find that two-dimensional persistent homology outperforms one-dimensional persistent homology in this application.

Yi-Hao Kao, Benjamin Van Roy, Daniel Rubin et al.

We consider the problem of learning a measure of distance among vectors in a feature space and propose a hybrid method that simultaneously learns from similarity ratings assigned to pairs of vectors and class labels assigned to individual vectors. Our method is based on a generative model in which class labels can provide information that is not encoded in feature vectors but yet relates to perceived similarity between objects. Experiments with synthetic data as well as a real medical image retrieval problem demonstrate that leveraging class labels through use of our method improves retrieval performance significantly.