2215 - MRI-Based Radiomic Features for Predicting Treatment Response in Unresectable Hepatocellular Carcinoma Receiving Sequential Transarterial Chemoembolization and Stereotactic Body Radiotherapy Followed by Immunotherapy: An Exploratory Study
Presenter(s)
S. Sheng1, K. S. K. Chan1, W. H. K. Chiu2, A. S. Lee3, N. S. M. Wong3, V. W. Y. Lee4, K. Man5, A. C. Y. Chan5, and C. L. Chiang1; 1Department of Clinical Oncology, School of Clinical Medicine, LKS Faculty of Medicine, The University of Hong Kong, Hong Kong, China, Hong Kong, Hong Kong, 2Department of Diagnostic Radiology, School of Clinical Medicine, LKS Faculty of Medicine, The University of Hong Kong, Hong Kong, China, Hong Kong, Hong Kong, 3Department of Clinical Oncology, Tuen Mun Hospital, Hong Kong, China, Hong Kong, Hong Kong, 4Medical Physics Unit, Department of Clinical Oncology, Tuen Mun Hospital, Hong Kong, China, Hong Kong, Hong Kong, 5Department of Surgery, School of Clinical Medicine, LKS Faculty of Medicine, The University of Hong Kong, Hong Kong, China, Hong Kong, Hong Kong
Purpose/Objective(s): Sequential transarterial chemoembolization and stereotactic body radiotherapy followed by immunotherapy (START-FIT) is a promising conversion therapy for unresectable hepatocellular carcinoma (HCC). Given the heterogeneity of treatment outcome, this study aims at investigating the feasibility of using baseline MRI radiomic features to predict treatment response.
Materials/Methods: Thirty-one patients treated with START-FIT were included (21 responders and 10 non-responders). Responders were patients achieving complete or partial response, while non-responders had stable or progressive disease per mRECIST criteria. Pre-treatment arterial and portal venous phase MRI scans were collected for semiautomatic 3D tumor segmentation using ITK-SNAP (v4.4.0) integrated with the nnInteractive deep learning framework. 1,962 radiomic features were then extracted via PyRadiomics, including shape features, first-order statistics, and texture features. Following constant features exclusion, Z-score normalization and redundancy removal (|r|>0.9), features were screened via univariate and LASSO regression to build a radiomics model. This was then integrated with a LASSO-based clinical model to create a combined model. Model performance was evaluated and compared using ROC curves with Leave-One-Out Cross-Validation and decision curve analysis (DCA).
Results: Twelve features were selected for radiomics model, 11 of which were wavelet-transformed, highlighting the predictive value of deep spatial frequencies. Notably, negative coefficient features are predominantly associated with tumor heterogeneity and non-uniform gray-level distribution, such as wavelet LHL gldm LargeDependenceHighGrayLevelEmphasis (ß= -1.11) in arterial phase, wavelet HHH glszm GrayLevelNonUniformity (ß= -0.55) and wavelet LLH glszm SizeZoneNonUniformityNormalized (ß= -0.41) in portal venous phase. The 12-feature radiomics model yielded an AUC of 0.89 (95% CI: 0.74 - 0.98), and the radiomics scores between responders (median 1.91; IQR, 1.35 - 2.58) and non-responders (median -1.77; IQR, -2.64 to -1.39) showed a significant difference (P < 0.001). The clinical model yielded an AUC of 0.72 (95% CI: 0.48 - 0.92), while the combined model demonstrated the highest predictive accuracy with an AUC of 0.91 (95% CI: 0.80 - 1.00), and DCA indicated that the combined model provides a substantial net benefit.
Conclusion: Pre-treatment MRI radiomic features are associated with heterogeneous response to START-FIT in HCC patients. This suggests radiomics could be a non-invasive tool for patient stratification, though further validation in larger cohorts is required to confirm its clinical application.