AI-based Recurrence Prediction in Invasive Lung Adenocarcinoma Using H&E-Stained Whole Slide Images

Introduction

Accurate prediction of recurrence risk after surgical resection in invasive lung adenocarcinoma (ILA) remains an unmet clinical need. Current prognostic methods face limitations such as interobserver variability and variable sensitivity in detecting residual disease. To address this, we developed an artificial intelligence (AI) based approach that leverages multiscale tissue analysis of hematoxylin and eosin (H&E) stained whole slide images (WSIs) to uncover morphological patterns linked to disease progression.

Methods

We utilized a dataset of 191 patients with stage I-III ILA who underwent surgical resection. H&E-stained WSIs of these patients were analyzed using a novel AI model to extract and integrate features across multiple cellular and tissue compartments. Submodels targeting distinct cell populations were developed and integrated to predict 5-year recurrence risk. Additionally, since we had patient-level ratios for seven key histological patterns provided by a thoracic pathologist, we trained another model to predict these ratios. Finally, we trained a model that incorporated patient demographics, clinical data, cellular features, histological features, and AI-predicted pattern ratios to predict recurrence.

Results

Our AI model achieved superior predictive performance (AUC: 78.22; accuracy: 72.49%; sensitivity: 62.40%; specificity: 78.66%) compared to conventional methods. Models focusing on single-cell populations, clinical data or pattern ratios only, showed moderate performance (AUC: 70.2–74.9), but their integration significantly enhanced performance. Our proposed model outperformed a baseline model trained on UNI2-h (AUC: 66.5; accuracy: 61.5%; sensitivity: 55.7%; specificity: 65.3%), highlighting the limitations of foundation models in capturing patterns critical for recurrence prediction. Our model’s ability to synthesize spatially resolved tissue architecture and cellular interactions contributed to its improved identification of high-risk patients.

Conclusions

These findings illustrate the potential of AI in identifying ILA patients at high risk of recurrence. By incorporating a broad spectrum of histological features and underutilized morphologic information, our model outperforms conventional methods and may guide adjuvant therapy decisions and follow-up strategies. Future work will focus on external validation, integration with molecular data, and developing interpretability frameworks for clinical adoption.