Uncertainty-Aware Convolutional Neural Network for Identifying Bilateral Opacities on Chest X-rays: A Tool to Aid Diagnosis of Acute Respiratory Distress Syndrome

• Type of Material: Article
• Authors: Mehak Arora, Emory University; Carolyn M. Davis, Emory University; Niraj R. Gowda, Emory University; Dennis G. Foster, Emory University; Angana Mondal, Emory University; Craig Coopersmith, Emory University; Rishikesan Kamaleswaran, Emory University
• Language: English
• Date: 2023-08-08
• Publisher: MDPI
• Publication Version: Final Published Version
• Copyright Statement: © 2023 by the authors.
• License: Creative Commons Attribution 4.0 International
• Final Published Version (URL): https://doi.org/10.3390%2Fbioengineering10080946
• Title of Journal or Parent Work: Bioengineering
• Volume: 10
• Issue: 8
• Start Page: 946
• Grant/Funding Information: R.K. and M.A. were supported by the National Institutes of Health (NIH) Award Number R01GM139967 and UL1TR002378. C.M.D. and C.M.C. were supported by NIH Award Number GM148217 and GM095442.
• Keywords: uncertainty modeling; acute respiratory distress syndrome; convolutional neural networks; medical image analysis; model calibration; chest X-ray classification; deep learning; computer-aided diagnosis
• Research Categories: Health Sciences, Radiology; Artificial Intelligence

Abstract

Acute Respiratory Distress Syndrome (ARDS) is a severe lung injury with high mortality, primarily characterized by bilateral pulmonary opacities on chest radiographs and hypoxemia. In this work, we trained a convolutional neural network (CNN) model that can reliably identify bilateral opacities on routine chest X-ray images of critically ill patients. We propose this model as a tool to generate predictive alerts for possible ARDS cases, enabling early diagnosis. Our team created a unique dataset of 7800 single-view chest-X-ray images labeled for the presence of bilateral or unilateral pulmonary opacities, or ‘equivocal’ images, by three blinded clinicians. We used a novel training technique that enables the CNN to explicitly predict the ‘equivocal’ class using an uncertainty-aware label smoothing loss. We achieved an Area under the Receiver Operating Characteristic Curve (AUROC) of 0.82 (95% CI: 0.80, 0.85), a precision of 0.75 (95% CI: 0.73, 0.78), and a sensitivity of 0.76 (95% CI: 0.73, 0.78) on the internal test set while achieving an (AUROC) of 0.84 (95% CI: 0.81, 0.86), a precision of 0.73 (95% CI: 0.63, 0.69), and a sensitivity of 0.73 (95% CI: 0.70, 0.75) on an external validation set. Further, our results show that this approach improves the model calibration and diagnostic odds ratio of the hypothesized alert tool, making it ideal for clinical decision support systems.