Detecting Acute Cellular Rejection in Lung Transplant Biopsies by Artificial Intelligence: A Novel Deep Learning Approach

Conference Paper

PURPOSE: Acute cellular rejection (ACR), and its relation to chronic lung allograft dysfunction (CLAD), is an important cause of morbidity and mortality in lung transplant patients. ACR occurs in more than one third of lung transplant recipients in the first year after transplantation; diagnosis largely rests on histopathologic evaluation of lung biopsy. The reproducibility of the diagnosis of ACR is quite variable, even among experienced transplant pathologists with published kappa values of interobserver agreement at 0.183 and 0.24-0.62. With digital pathology and artificial intelligence (AI) technology, it may be possible to aid pathologists in the diagnosis of ACR. Herein we examine, for the first time, if an AI machine learning algorithm can reliably distinguish the vascular component of rejection in lung transplant biopsies from normal lung. METHODS: At a single high-volume lung transplant center, annotations were completed by board-certified lung transplant pathologists from hematoxylin and eosin stained slides scanned using a Leica Aperio AT2 digital whole slide scanner at 40X magnification. Annotations were imported into a Python 3.7 development environment using the open-source 'Openslide' library. Keras with a tensorflow backend was used for AI modeling and training. All training was performed on a Windows 10 based PC with an Intel Core I-9 processor, 64gb of ram, and Nvidia GPU. RESULTS: A total of 3,349 annotations (2580 regions of normal, 769 lesions of A1/A2 rejection) were completed. For the training set, 614 A1/A2 lesions and 2064 regions of normal were evaluated. For the validation set, 155 A1/A2 lesions and 156 regions of normal were used to test the AI algorithm's accuracy. The algorithm distinguished the vascular component of ACR from normal alveolated lung tissue with 95% validation accuracy. CONCLUSION: To our knowledge, our study is the first to provide evidence that an AI machine algorithm can reach acceptable, and possibly surpass, performance of current diagnostic standards of identifying ACR in lung transplant patients. Future studies should include multi-institutional validation testing.

Full Text

Duke Authors

Cited Authors

  • Davis, H; Glass, C; Davis, RC; Glass, M; Pavlisko, EN

Published Date

  • April 1, 2020

Published In

Volume / Issue

  • 39 / 4

Start / End Page

  • S501 - S502

Electronic International Standard Serial Number (EISSN)

  • 1557-3117

Digital Object Identifier (DOI)

  • 10.1016/j.healun.2020.01.100

Citation Source

  • Scopus