Whole‐slide margin control through deep learning in Mohs micrographic surgery for basal cell carcinoma

Background Basal cell carcinoma (BCC) is the most common type of skin cancer with incidence rates rising each year. Mohs micrographic surgery (MMS) is most often chosen as treatment for BCC on the face for which each frozen section has to be histologically analysed to ensure complete tumor removal....

Full description

Saved in:
Bibliographic Details
Published in:Experimental dermatology 2021-05, Vol.30 (5), p.733-738
Main Authors: Zon, Mike C.M., Waa, José D., Veta, Mitko, Krekels, Gertruud A.M.
Format: Article
Language:English
Subjects:
Basal cell carcinoma
Carcinoma, Basal Cell - pathology
Carcinoma, Basal Cell - surgery
Deep Learning
dermatopathology
digital pathology
Humans
Margins of Excision
Micrography
Mohs micrographic surgery
Mohs Surgery
Neoplasm Invasiveness
Neural networks
Segmentation
Skin cancer
Skin Neoplasms - pathology
Skin Neoplasms - surgery
Surgery
Citations: Items that this one cites
Items that cite this one
Online Access:Get full text
Tags: Add Tag
No Tags, Be the first to tag this record!
Description
Summary:Background Basal cell carcinoma (BCC) is the most common type of skin cancer with incidence rates rising each year. Mohs micrographic surgery (MMS) is most often chosen as treatment for BCC on the face for which each frozen section has to be histologically analysed to ensure complete tumor removal. This causes a heavy burden on health economics. Objectives To develop and evaluate a deep learning model for the automated detection of BCC‐negative slides and classification of BCC in histopathology slides of MMS based on whole‐slide image (WSI). Methods Two deep learning models were developed on the basis of 171 digitized H&E frozen slides from 70 different patients. The first model had a U‐Net architecture and was used for the segmentation of BCC. A subsequent convolutional neural network used the segmentation to classify the whole slide as BCC or BCC‐negative. Results Quantitative evaluation over manually labelled ground truth data resulted in a Dice score of 0.66 for the segmentation of BCC and an area under the receiver operating characteristic curve (AUC) of 0.90 for the slide‐level classification. Conclusions This study demonstrates that through WSIs deep learning models may be a feasible option to improve the clinical workflow and reduce costs in histological analysis of BCC in MMS.
ISSN:0906-6705
1600-0625
DOI:10.1111/exd.14306