Deep learning-based pRb subtyping of pulmonary large cell neuroendocrine carcinoma on small hematoxylin and eosin-stained specimens.

Abstract

Molecular subtyping of pulmonary large cell neuroendocrine carcinoma (LCNEC) based on retinoblastoma protein (pRb) expression may influence systemic treatment decisions. Current histomorphological assessments of hematoxylin and eosin (H&E)-stained tissue samples cannot reliably differentiate LCNEC molecular subtypes. This study explores the potential of deep learning (DL) to identify histological patterns that distinguish these subtypes, by developing a custom convolutional neural network to predict the binary expression of pRb in small LCNEC tissue samples. Our model was trained, cross-validated, and tested on tissue micro-array cores from 143 resection specimens and biopsies from 21 additional patients, with corresponding immunohistochemical pRb status. The best-performing DL model achieved a patient-wise balanced accuracy of 0.75 and a ROC-AUC of 0.77 when tested on biopsies, significantly outperforming the H&E-based subtype classification by lung pathologists. Explainable artificial intelligence techniques further highlighted coarse chromatin patterns and distinct nucleoli as distinguishing features for pRb retained status. Meanwhile pRb lost cases were characterized by limited cytoplasm and morphological similarities with small cell lung cancer. These findings suggest that DL analysis of small histopathology samples could ultimately replace immunohistochemical pRb testing. Such a development may assist in guiding chemotherapy decisions, particularly in metastatic cases.