A deep learning model to predict objective response to TACE and TKI-based therapy in HBV-related uHCC

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Hepatocellular Carcinoma Research

Background:

• Transarterial chemoembolization (TACE) and tyrosine kinase inhibitor (TKI) combination therapy show promise for unresectable hepatocellular carcinoma (uHCC).
• Significant inter-patient heterogeneity in treatment response necessitates reliable biomarkers for personalized management.
• Hepatitis B virus (HBV) infection is a common driver for uHCC, influencing treatment efficacy.

Purpose of the Study:

• To develop and validate a deep learning (DL) model for predicting objective response and progression-free survival (PFS) in HBV-related uHCC patients undergoing TACE and TKI therapy.
• To compare the performance of the DL model against a clinical model (C-Model) and a machine learning radiomics model (ML-Model).
• To enhance model interpretability using Grad-CAM visualization to understand spatial patterns of tumor response.

Main Methods:

• Retrospective analysis of 243 HBV-related uHCC patients, divided into training and testing datasets.
• Development of three predictive models: C-Model (clinical data), ML-Model (1,479 CT radiomic features), and DL-Model (ResNet-50 architecture).
• Performance evaluation using Area Under the Curve (AUC) for response prediction and Kaplan-Meier analysis for PFS, with Grad-CAM for interpretability.

Main Results:

• The DL-Model achieved a superior AUC of 0.851 in the test set, significantly outperforming the C-Model (AUC=0.586) and ML-Model (AUC=0.709).
• The DL-Model was the only framework capable of robust prognostic stratification, identifying predicted responders with significantly prolonged PFS (P=0.011).
• Grad-CAM analysis revealed distinct spatial activation patterns: focal/centralized in responders versus multifocal/peripheral in non-responders.

Conclusions:

• The developed DL-Model offers a reliable and interpretable tool for predicting treatment response and PFS in uHCC patients receiving TACE and TKI-based therapy.
• The DL-Model's ability to stratify patients based on predicted response facilitates personalized treatment adjustments.
• Grad-CAM visualization provides valuable spatial insights into tumor heterogeneity, aiding clinical decision-making for uHCC management.