Deep learning-based histomorphological subtyping and risk stratification of small cell lung cancer from hematoxylin and eosin-stained whole slide images.

BACKGROUND

Accurate subtyping and risk stratification are imperative for prognostication and clinical decision-making in small cell lung cancer (SCLC). However, traditional molecular subtyping is resource-intensive and challenging to translate into clinical practice.

METHODS

A total of 517 SCLC patients and their corresponding hematoxylin and eosin (H&E)-stained whole slide images (WSIs) from three independent medical institutions were analyzed. A hybrid clustering-based unsupervised deep representation learning model was developed to identify histomorphological phenotypes (HIPO) and characterize tumor ecosystem diversity. Consensus clustering and a deep learning-based stratification system were used to define histomorphological subtypes (HIPOS) based on patient-level HIPO features. Survival analysis and Cox proportional hazards regression models were used to assess the clinical significance of HIPOS. An integrated analysis of pathomics, proteomics, and immunohistochemistry was conducted to explore the biological and microenvironmental correlates of HIPOS.

RESULTS

We performed histomorphological phenotyping of SCLC using unsupervised deep representation learning from WSIs and identified 15 HIPOs. Unsupervised clustering of HIPO profiles stratified SCLCs into two reproducible image-based subtypes: HIPOS-I and HIPOS-II. Patients in the HIPOS-I group had better overall survival and disease-free survival compared to those in HIPOS-II, independent of clinical features and molecular subtypes. Multimodal analyses revealed that HIPOS-I tumors were characterized by enriched immune infiltration and immune activation, whereas HIPOS-II tumors displayed increased fibrosis, cellular pleomorphism, and dysregulated oxidative metabolism. Additionally, we developed a simplified deep-learning model to predict HIPOS subtypes to enhance clinical applications and validated the prognostic value of these subtypes in independent cohorts.

CONCLUSIONS

This study demonstrates the potential of a deep learning-based histomorphological subtyping system to improve patient stratification and prognosis prediction in SCLC. The HIPOS offers a promising and clinically applicable tool for personalized management using routine H&E-stained WSIs.