AI Improves Tumor Proportion Scoring in Lung Cancer

Accurately determining the tumor proportion score (TPS) in non-small cell lung cancer (NSCLC) is crucial for patient treatment and prognosis. However, this process is labor-intensive, requires highly specialized experts, and involves manually annotating numerous slides. Existing automated methods, particularly those based on multiple instance learning (MIL), often struggle with images that contain no tumor cells. Researchers have developed a new end-to-end framework to address these challenges. It comprises two main components: an embedding network that identifies features from individual image patches and a MIL model that combines these features. Crucially, the MIL model predicts parameters for a zero-inflated beta (ZIBeta) probability distribution, which represents the overall TPS for an entire slide. This distribution-based approach allows the system to handle images with varying tumor proportions, including those with zero tumor cells, more effectively than previous methods. By using only slide-level TPS scores for training, the framework demonstrates improved prediction accuracy and enhanced explainability compared to traditional linear and ridge regression baselines.