New research on using deep learning to automatically score PD-L1 expression in NSCLC published in Biomolecules and Biomedicine

‍As published in Biomolecules and Biomedicine Advance online (2025)

A novel deep learning framework for automatic scoring of PD-L1 expression in non-small cell lung cancer

• Saidul Kabir Department of Electrical and Electronic Engineering, University of Dhaka, Dhaka, Bangladesh
• Muhammad E. H. Chowdhury Department of Electrical Engineering, Qatar University, Doha, Qatar
• Rusab Sarmun Department of Electrical and Electronic Engineering, University of Dhaka, Dhaka, Bangladesh
• Semir Vranić College of Medicine, QU Health, Qatar University, Doha, Qatar https://orcid.org/0000-0001-9743-7265
• Rafif Mahmood Al Saady College of Medicine, QU Health, Qatar University, Doha, Qatar
• Inga Rose Reference Medicine, Phoenix, Arizona, United States of America
• Zoran Gatalica Reference Medicine, Phoenix, Arizona, United States of America

Abstract

A critical predictive marker for anti-PD-1/PD-L1 therapy is programmed death-ligand 1 (PD-L1) expression, assessed by immunohistochemistry (IHC). This paper explores a novel automated framework using deep learning to accurately evaluate PD-L1 expression from whole slide images (WSIs) of non-small cell lung cancer (NSCLC), aiming to improve the precision and consistency of Tumor Proportion Score (TPS) evaluation, which is essential for determining patient eligibility for immunotherapy. Automating TPS evaluation can enhance accuracy and consistency while reducing pathologists' workload. The proposed automated framework encompasses three stages: identifying tumor patches, segmenting tumor areas, and detecting cell nuclei within these areas, followed by estimating the TPS based on the ratio of positively stained to total viable tumor cells. This study utilized a Reference Medicine (Phoenix, Arizona) dataset containing 66 NSCLC tissue samples, adopting a hybrid human-machine approach for annotating extensive WSIs. Patches of size 1000x1000 pixels were generated to train classification models such as EfficientNet, Inception, and Vision Transformer models. Additionally, segmentation performance was evaluated across various UNet and DeepLabV3 architectures, and the pre-trained StarDist model was employed for nuclei detection, replacing traditional watershed techniques. PD-L1 expression was categorized into three levels based on TPS: negative expression (TPS < 1%), low expression (TPS 1-49%), and high expression (TPS ≥ 50%). The Vision Transformer-based model excelled in classification, achieving an F1-score of 97.54%, while the modified DeepLabV3+ model led in segmentation, attaining a Dice Similarity Coefficient of 83.47%. The TPS predicted by the framework closely correlated with the pathologist's TPS at 0.9635, and the framework's three-level classification F1-score was 93.89%. The proposed deep learning framework for automatically evaluating the TPS of PD-L1 expression in NSCLC demonstrated promising performance. This framework presents a potential tool that could produce clinically significant results more efficiently and cost-effectively.

Download the manuscript from Biomolecules and Biomedicine.