Deep Semi Supervised Generative Learning for Automated Tumor Proportion Scoring on NSCLC Tissue Needle Biopsies

Area of Science:

• Oncology
• Computational Pathology
• Biomarker Discovery

Background:

• PD-L1 expression is a critical biomarker for predicting response to anti PD-1/PD-L1 therapies in Non-Small-Cell-Lung-Cancer (NSCLC).
• Current visual estimation of PD-L1 (tumor proportional scoring or TPS) by pathologists is subjective and prone to variability, especially with challenging biopsy samples.
• Objective and automated quantification of PD-L1 is needed to improve treatment selection accuracy.

Purpose of the Study:

• To develop and validate a novel deep learning solution for automated and objective scoring of PD-L1 expression in NSCLC needle biopsies.
• To address challenges of limited tissue and manual annotation in biopsy images using semi-supervised learning.
• To compare the performance of the automated method against pathologist visual scoring.

Main Methods:

• Development of a deep learning model for automated PD-L1 scoring.
• Utilized semi-supervised learning approaches to train the model with limited manual annotations.
• Consolidated manual annotations and pathologist-derived TPS scores for training and evaluation.
• Quantitative evaluation using concordance measures on an independent set of slides.

Main Results:

• The proposed deep learning method achieved automated and objective scoring of PD-L1 expression.
• Semi-supervised learning was effective in handling limited tissue and annotation data.
• The automated scoring demonstrated concordance with visual scoring by multiple pathologists.
• The method ensures repeatability and objectivity, overcoming limitations of manual assessment.

Conclusions:

• A novel deep learning approach provides an automated and objective method for PD-L1 scoring in NSCLC biopsies.
• This technology has the potential to enhance the accuracy and consistency of biomarker assessment for immunotherapy selection.
• The use of semi-supervised learning is a viable strategy for developing computational pathology tools with limited annotated data.