AI Algorithm for Lung Adenocarcinoma Pattern Quantification (PATQUANT): International Validation and Advanced Risk Stratification Superior to Conventional Grading

Abstract

The morphological patterns of lung adenocarcinoma (LUAD) are recognized for their prognostic significance, with ongoing debate regarding the optimal grading strategy. This study aimed to develop a clinical-grade, fully quantitative, and automated tool for pattern classification/quantification (PATQUANT), to evaluate existing grading strategies, and determine the optimal grading system. PATQUANT was trained on a high-quality dataset, manually annotated by expert pathologists. Several independent test datasets and 13 expert pathologists were involved in validation. Five large, multinational cohorts of resectable LUAD (patient n = 1120) were analyzed concerning prognostic value. PATQUANT demonstrated excellent pattern segmentation/classification accuracy and outperformed 8 out of 13 pathologists. The prognostic study revealed a distinct prognostic profile for the complex glandular pattern. While all contemporary grading systems had prognostic value, the predominant pattern-based and simplified IASLC systems were superior. We propose and validate two new, fully explainable grading principles, providing fine-grained, statistically independent patient risk stratification. We developed a fully automated, robust AI tool for pattern analysis/quantification that surpasses the performance of experienced pathologists. Additionally, we demonstrate the excellent prognostic capabilities of two new grading approaches that outperform traditional grading methods. We make our extensive agreement dataset publicly available to advance the developments in the field.