Workload impact of automated image analysis and deep learning on manual screening of sentinel node biopsies in breast cancer

J. Thagaard, M. Kristensson
DTU Compute/Visiopharm, Lyngby, Denmark

Objective: Our aim of this study is to evaluate whether digital image analysis (DIA) and deep learning (DL) can decrease the workload for the examining pathologist when screening sentinel node biopsies (SLNB) in breast cancer (BC) - without compromising the diagnostic accuracy.
Method: From a cohort of 135 patients with BC receiving surgery, SLNB were collected from Dept. of Pathology Rigshospitalet, Odense and Slagelse, Denmark according to national and international guidelines. Tissue samples were submitted for frozen section procedure, where serial sections were stained locally with immunohistochemistry (IHC) and hematoxylin and eosin (H&E). Stained sections were digitized and analyzed using a DIA algorithm for IHC cytokeratin (CK) and a DL-based algorithm for H&E.
Results: Conventional microscopy was used as golden standard and compared with DIA and DL. The IHC-CK algorithm demonstrated a sensitivity of 100 % (i.e. no false negative slides were observed). On average, the workload could have been decreased by 58.2 % by using DIA as a screening tool. We aim to include results from H&E DL algorithm in the future. Conclusion: Our proposed IHC-CK algorithm is an ideal screening tool for SLNB, and implementation of DIA has already shown a decrease in workload for examining pathologists by over 50 %.