Development and validation of a supervised deep learning algorithm for automated whole‐slide programmed death‐ligand 1 tumour proportion score assessment in non‐small cell lung cancer. Issue 4 (16th November 2021)

• Record Type: Journal Article
• Authors: Hondelink, Liesbeth M; Hüyük, Melek; Postmus, Pieter E; Smit, Vincent T H B M; Blom, Sami; von der Thüsen, Jan H; Cohen, Danielle
• Is Part Of: Histopathology. Volume 80:Issue 4(2022)
• Page Start: 635
• Page End: 647
• Publication Date: 2021-11-16
• Subjects: artificial intelligence -- computational pathology -- immunotherapy -- non‐small cell lung cancer -- programmed death‐ligand 1; Histology, Pathological -- Periodicals; 611.018
• DOI: 10.1111/his.14571
• Languages: English
• ISSNs: 0309-0167

Abstract

Abstract : Aims: Immunohistochemical programmed death‐ligand 1 (PD‐L1) staining to predict responsiveness to immunotherapy in patients with advanced non‐small cell lung cancer (NSCLC) has several drawbacks: a robust gold standard is lacking, and there is substantial interobserver and intraobserver variance, with up to 20% discordance around cutoff points. The aim of this study was to develop a new deep learning‐based PD‐L1 tumour proportion score (TPS) algorithm, trained and validated on a routine diagnostic dataset of digitised PD‐L1 (22C3, laboratory‐developed test)‐stained samples. Methods and results: We designed a fully supervised deep learning algorithm for whole‐slide PD‐L1 assessment, consisting of four sequential convolutional neural networks (CNNs), using aiforia create software. We included 199 whole slide images (WSIs) of 'routine diagnostic' histology samples from stage IV NSCLC patients, and trained the algorithm by using a training set of 60 representative cases. We validated the algorithm by comparing the algorithm TPS with the reference score in a held‐out validation set. The algorithm had similar concordance with the reference score (79%) as the pathologists had with one another (75%). The intraclass coefficient was 0.96 and Cohen's κ coefficient was 0.69 for the algorithm. Around the 1% and 50% cutoff points, concordance was also similar between pathologists and the algorithm. Conclusions: We designed a new, deep learning‐based PD‐L1 TPS algorithm that is similarly able to assess PD‐L1 expression in daily routine diagnostic cases as pathologists. Successful validation on routine diagnostic WSIs and detailed visual feedback show that this algorithm meets the requirements for functioning as a 'scoring assistant'. Abstract :