Non-Invasive Measurement Using Deep Learning Algorithm Based on Multi-Source Features Fusion to Predict PD-L1...

Area of Science:

• Oncology
• Radiology
• Artificial Intelligence

Background:

• Programmed death-ligand 1 (PD-L1) is a key biomarker for lung cancer immunotherapy.
• Assessing PD-L1 expression via tumor proportion score (TPS) is challenging due to invasive sampling and heterogeneity.
• There is a need for non-invasive methods to measure PD-L1 expression signature (ES).

Purpose of the Study:

• To develop and validate an artificial intelligence (AI) system for non-invasive PD-L1 expression signature (ES) measurement in non-small cell lung cancer (NSCLC).
• To evaluate the AI system's ability to predict PD-L1 ES levels and patient survival outcomes.
• To assess the added value of AI combined with clinical factors for treatment decision-making.

Main Methods:

• Developed a deep learning (DL) system incorporating radiomics and combination models using computed tomography (CT) images from 1,135 NSCLC patients.
• Utilized a 3D ResNet for feature extraction and a specialized classifier for PD-L1 ES prediction.
• Employed a Cox proportional-hazards model with clinical factors and PD-L1 ES for survival analysis.

Main Results:

• The combination AI model achieved high performance in predicting PD-L1 ES categories (<1%, 1-49%, ≥50%) with AUCs ranging from 0.933 to 0.950 in the validation cohort.
• The AI model incorporating multi-source features demonstrated superior overall survival prediction (C-index: 0.89) compared to clinical models alone (C-index: 0.86).

Conclusions:

• A non-invasive deep learning-based AI system can accurately assess PD-L1 expression and predict survival outcomes in NSCLC.
• Combining DL models with clinical characteristics enhances prediction capabilities, supporting clinical treatment decisions.
• This AI approach offers a promising non-invasive tool for managing NSCLC patients undergoing immunotherapy.