We focus on the problem of producing well-calibrated out-of-distribution (OOD) detectors, in order to enable safe deployment of medical image classifiers. Motivated by the difficulty of curating suitable calibration datasets, synthetic augmentations have become highly prevalent for inlier/outlier specification. While there have been rapid advances in data augmentation techniques, this paper makes a striking finding that the space in which the inliers and outliers are synthesized, in addition to the type of augmentation, plays a critical role in calibrating OOD detectors. Using the popular energy-based OOD detection framework, we find that the optimal protocol is to synthesize latent-space inliers along with diverse pixel-space outliers. Based on empirical studies with multiple medical imaging benchmarks, we demonstrate that our approach consistently leads to superior OOD detection ($15\% - 35\%$ in AUROC) over the state-of-the-art in a variety of open-set recognition settings.

With increased interest in adopting AI methods for clinical diagnosis, a vital step towards safe deployment of such tools is to ensure that the models not only produce accurate predictions but also do not generalize to data regimes where the training data provide no meaningful evidence. Existing approaches for ensuring the distribution of model predictions to be similar to that of the true distribution rely on explicit uncertainty estimators that are inherently hard to calibrate. In this paper, we propose to train a loss estimator alongside the predictive model, using a contrastive training objective, to directly estimate the prediction uncertainties. Interestingly, we find that, in addition to producing well-calibrated uncertainties, this approach improves the generalization behavior of the predictor. Using a dermatology use-case, we show the impact of loss estimators on model generalization, in terms of both its fidelity on in-distribution data and its ability to detect out of distribution samples or new classes unseen during training.

Pulmonary embolisms (PE) are known to be one of the leading causes for cardiac-related mortality. Due to inherent variabilities in how PE manifests and the cumbersome nature of manual diagnosis, there is growing interest in leveraging AI tools for detecting PE. In this paper, we build a two-stage detection pipeline that is accurate, computationally efficient, robust to variations in PE types and kernels used for CT reconstruction, and most importantly, does not require dense annotations. Given the challenges in acquiring expert annotations in large-scale datasets, our approach produces state-of-the-art results with very sparse emboli contours (at 10mm slice spacing), while using models with significantly lower number of parameters. We achieve AUC scores of 0.94 on the validation set and 0.85 on the test set of highly severe PEs. Using a large, real-world dataset characterized by complex PE types and patients from multiple hospitals, we present an elaborate empirical study and provide guidelines for designing highly generalizable pipelines.

In this working notes paper, we describe IBM Research AI (Almaden) team's participation in the ImageCLEF 2019 VQA-Med competition. The challenge consists of four question-answering tasks based on radiology images. The diversity of imaging modalities, organs and disease types combined with a small imbalanced training set made this a highly complex problem. To overcome these difficulties, we implemented a modular pipeline architecture that utilized transfer learning and multi-task learning. Our findings led to the development of a novel model called Supporting Facts Network (SFN). The main idea behind SFN is to cross-utilize information from upstream tasks to improve the accuracy on harder downstream ones. This approach significantly improved the scores achieved in the validation set (18 point improvement in F-1 score). Finally, we submitted four runs to the competition and were ranked seventh.

Acceleration of machine learning research in healthcare is challenged by lack of large annotated and balanced datasets. Furthermore, dealing with measurement inaccuracies and exploiting unsupervised data are considered to be central to improving existing solutions. In particular, a primary objective in predictive modeling is to generalize well to both unseen variations within the observed classes, and unseen classes. In this work, we consider such a challenging problem in machine learning driven diagnosis: detecting a gamut of cardiovascular conditions (e.g. infarction, dysrhythmia etc.) from limited channel ECG measurements. Though deep neural networks have achieved unprecedented success in predictive modeling, they rely solely on discriminative models that can generalize poorly to unseen classes. We argue that unsupervised learning can be utilized to construct effective latent spaces that facilitate better generalization. This work extensively compares the generalization of our proposed approach against a state-of-the-art deep learning solution. Our results show significant improvements in F1-scores.

The hypothesis that computational models can be reliable enough to be adopted in prognosis and patient care is revolutionizing healthcare. Deep learning, in particular, has been a game changer in building predictive models, thereby leading to community-wide data curation efforts. However, due to the inherent variabilities in population characteristics and biological systems, these models are often biased to the training datasets. This can be limiting when models are deployed in new environments, particularly when there are systematic domain shifts not known a priori. In this paper, we formalize these challenges by emulating a large class of domain shifts that can occur in clinical settings, and argue that evaluating the behavior of predictive models in light of those shifts is an effective way of quantifying the reliability of clinical models. More specifically, we develop an approach for building challenging scenarios, based on analysis of \textit{disease landscapes}, and utilize unsupervised domain adaptation to compensate for the domain shifts. Using the openly available MIMIC-III EHR dataset for phenotyping, we generate a large class of scenarios and evaluate the ability of deep clinical models in those cases. For the first time, our work sheds light into data regimes where deep clinical models can fail to generalize, due to significant changes in the disease landscapes between the source and target landscapes. This study emphasizes the need for sophisticated evaluation mechanisms driven by real-world domain shifts to build effective AI solutions for healthcare.

Processing temporal sequences is central to a variety of applications in health care, and in particular multi-channel Electrocardiogram (ECG) is a highly prevalent diagnostic modality that relies on robust sequence modeling. While Recurrent Neural Networks (RNNs) have led to significant advances in automated diagnosis with time-series data, they perform poorly when models are trained using a limited set of channels. A crucial limitation of existing solutions is that they rely solely on discriminative models, which tend to generalize poorly in such scenarios. In order to combat this limitation, we develop a generative modeling approach to limited channel ECG classification. This approach first uses a Seq2Seq model to implicitly generate the missing channel information, and then uses the latent representation to perform the actual supervisory task. This decoupling enables the use of unsupervised data and also provides highly robust metric spaces for subsequent discriminative learning. Our experiments with the Physionet dataset clearly evidence the effectiveness of our approach over standard RNNs in disease prediction.

With widespread adoption of electronic health records, there is an increased emphasis for predictive models that can effectively deal with clinical time-series data. Powered by Recurrent Neural Network (RNN) architectures with Long Short-Term Memory (LSTM) units, deep neural networks have achieved state-of-the-art results in several clinical prediction tasks. Despite the success of RNN, its sequential nature prohibits parallelized computing, thus making it inefficient particularly when processing long sequences. Recently, architectures which are based solely on attention mechanisms have shown remarkable success in transduction tasks in NLP, while being computationally superior. In this paper, for the first time, we utilize attention models for clinical time-series modeling, thereby dispensing recurrence entirely. We develop the SAnD (Simply Attend and Diagnose) architecture, which employs a masked, self-attention mechanism, and uses positional encoding and dense interpolation strategies for incorporating temporal order. Furthermore, we develop a multi-task variant of SAnD to jointly infer models with multiple diagnosis tasks. Using the recent MIMIC-III benchmark datasets, we demonstrate that the proposed approach achieves state-of-the-art performance in all tasks, outperforming LSTM models and classical baselines with hand-engineered features.

With widespread adoption of electronic health records, there is an increased emphasis for predictive models that can effectively deal with clinical time-series data. Powered by Recurrent Neural Network (RNN) architectures with Long Short-Term Memory (LSTM) units, deep neural networks have achieved state-of-the-art results in several clinical prediction tasks. Despite the success of RNNs, its sequential nature prohibits parallelized computing, thus making it inefficient particularly when processing long sequences. Recently, architectures which are based solely on attention mechanisms have shown remarkable success in transduction tasks in NLP, while being computationally superior. In this paper, for the first time, we utilize attention models for clinical time-series modeling, thereby dispensing recurrence entirely. We develop the \textit{SAnD} (Simply Attend and Diagnose) architecture, which employs a masked, self-attention mechanism, and uses positional encoding and dense interpolation strategies for incorporating temporal order. Furthermore, we develop a multi-task variant of \textit{SAnD} to jointly infer models with multiple diagnosis tasks. Using the recent MIMIC-III benchmark datasets, we demonstrate that the proposed approach achieves state-of-the-art performance in all tasks, outperforming LSTM models and classical baselines with hand-engineered features.