BraMARS: An Interpretable Histopathology-Driven Deep Learning Model for Brain Metastasis Risk Stratification in...

Area of Science:

• Oncology
• Artificial Intelligence
• Pathology

Background:

• Brain metastasis (BM) is a primary cause of death in limited-stage small-cell lung cancer (LS-SCLC).
• Prophylactic cranial irradiation (PCI) is used to prevent BM but has neurotoxic side effects and lacks personalized risk assessment.

Purpose of the Study:

• To develop an explainable deep learning model (BraMARS) for estimating future BM risk in resected LS-SCLC patients.
• To assess the model's performance and its potential to guide individualized treatment decisions.

Main Methods:

• Developed BraMARS, an AI model analyzing H&E-stained whole-slide images of resected LS-SCLC.
• Validated the model across independent patient cohorts.
• Performed histopathologic attribution and proteomic analyses to understand model predictions.

Main Results:

• BraMARS showed strong discriminatory performance (AUCs 0.738–0.944) and stratified patients by BM risk.
• High-risk patients had significantly worse survival outcomes.
• Simulations indicated potential to reduce PCI in 19.3% of low-risk patients while improving high-risk identification by 84.4%.

Conclusions:

• BraMARS offers a biologically interpretable, histopathology-based method for predicting BM risk in LS-SCLC.
• The model can support personalized intracranial risk assessment and surveillance strategies.
• Findings may guide future BM prevention strategies.