Date Published: April 3, 2018
Publisher: Public Library of Science
Author(s): Jeffrey J. Nirschl, Andrew Janowczyk, Eliot G. Peyster, Renee Frank, Kenneth B. Margulies, Michael D. Feldman, Anant Madabhushi, Alison Marsden.
Over 26 million people worldwide suffer from heart failure annually. When the cause of heart failure cannot be identified, endomyocardial biopsy (EMB) represents the gold-standard for the evaluation of disease. However, manual EMB interpretation has high inter-rater variability. Deep convolutional neural networks (CNNs) have been successfully applied to detect cancer, diabetic retinopathy, and dermatologic lesions from images. In this study, we develop a CNN classifier to detect clinical heart failure from H&E stained whole-slide images from a total of 209 patients, 104 patients were used for training and the remaining 105 patients for independent testing. The CNN was able to identify patients with heart failure or severe pathology with a 99% sensitivity and 94% specificity on the test set, outperforming conventional feature-engineering approaches. Importantly, the CNN outperformed two expert pathologists by nearly 20%. Our results suggest that deep learning analytics of EMB can be used to predict cardiac outcome.
Cardiovascular diseases are the leading cause of death globally and the leading cause of hospital admissions in the United States and Europe . More than 26 million people worldwide suffer from heart failure annually and about half of these patients die within five years [2, 3]. Heart failure is a serious, progressive clinical syndrome where impaired ventricular function results in inadequate systemic perfusion. The diagnosis of heart failure usually relies on clinical history, physical exam, basic lab tests, and imaging . However, when the cause of heart failure is unidentified, endomyocardial biopsy (EMB) represents the gold standard for the evaluation and grading of heart disease . The primary concern with the manual interpretation of EMB is the relatively high inter-rater variability  and limited clinical indications [5, 7]. Automated analysis and grading of cardiac histopathology can serve as an objective second read to reduce variability.
In this study, we developed a CNN classifier to detect clinical heart failure from cardiac histopathology. Previous studies that have applied deep learning to digital pathology have used CNNs to generate pixel-level cancer likelihood maps [14, 18] or segment relevant biological structures (e.g. glands, mitoses, nuclei, etc.) that are used as features for subsequent classification [27, 37]. However, our CNN directly transforms an image into a probability of a patient-level diagnosis, which is similar to recent approaches that have applied CNNs to diagnose referable diabetic retinopathy and skin cancer [29, 30].