Two-Stage Deep Learning Framework for Histopathological Subtyping of Non-Small Cell Lung Carcinoma in Small Biopsies

Artificial intelligence (AI)-based technologies have the potential to improve diagnostic accuracy while reducing tissue loss in small biopsies. In non-small-cell lung carcinoma (NSCLC), small biopsies often limit ancillary testing. We aimed to develop a deep learning-based framework for automated histopathological classification of NSCLC subtypes. Archived hematoxylin and eosin-stained whole-slide images from 341 NSCLC and benign specimens were digitized and annotated by two expert pathologists. A two-stage Dense Net121-based cascade was employed: Stage 1 differentiated not otherwise specified (NOS) from not-NOS specimens; Stage 2 classified not-NOS into adenocarcinoma, squamous cell carcinoma, or non-cancerous tissue. Patient-level predictions were derived via majority voting. Performance was evaluated using accuracy, precision, recall, and F1-score. In Stage 1, NOS versus not-NOS classification achieved 0.92 patient-level accuracy. In Stage 2, accuracy for adenocarcinoma, squamous cell carcinoma, and non-cancer was 0.94, with perfect classification of non-cancer patients (precision, recall, and F1 = 1.00). In the final step, the combined analysis demonstrated an overall accuracy of 0.88, with recall rates of 1.00 for adenocarcinoma and non-cancer, 0.83 for squamous cell carcinoma, and 0.67 for NOS. The proposed two-stage model demonstrated strong performance across NSCLC subtypes, including the diagnostically challenging NOS group. This approach may reduce reliance on ancillary tests, accelerate diagnosis, and optimize tissue preservation.