98 - Radiomics-Based Machine Learning to Predict Post-TACE Outcomes Using Lipiodol Staining Patterns in Hepatocellular Carcinoma

Purpose: To evaluate whether radiomics features derived from lipiodol staining patterns on post-TACE CT imaging can be used in conjunction with patient-level data to accurately predict clinically significant outcomes in hepatocellular carcinoma (HCC), including 6-month Milan criteria status, target lesion mRECIST response, progression-free survival (PFS), and overall survival (OS).

Materials and Methods: Two hundred patients with biopsy-confirmed HCC treated with selective transarterial chemoembolization (TACE) at a tertiary care center were retrospectively included. Non-contrast CT imaging within 30 days post-TACE was used to assess lipiodol staining. Lesions were segmented using the Segment Anything 2 (SAM2) model, enabling semi-automated extraction of lipiodol deposition regions. Radiomic features were derived from each segmented lesion using the PyRadiomics library, including shape, intensity, and texture-based metrics. These imaging features were integrated with clinical variables such as age, sex, Child-Pugh class, MELD score, BCLC stage, and systemic therapy status at the time of TACE. A machine learning model was trained using AutoGluon’s ensemble framework. The cohort was split into 160 patients for training and 40 for independent testing. Primary outcomes included binary Milan criteria status and mRECIST response at 6 months, as well as continuous outcomes for PFS and OS. Model performance was evaluated using AUC, sensitivity (SE), specificity (SP), positive predictive value (PPV), negative predictive value (NPV), accuracy, F1 score, and R² for survival.

Results: The final AutoML model was an ensemble model with a NeuralNetFastAI model in the first layer:

• 
  AUC (Milan) = 0.93
  
  


• 
  SE = 89.5%, SP = 85.7%, PPV = 85.0%, NPV = 90.0%, Accuracy = 87.5%, F1 = 87.2%
  
  


• 
  AUC (mRECIST) = 0.81
  
  


• 
  R² (PFS) = 0.83, R² (OS) = 0.79
  
  

Conclusion: Post-TACE lipiodol staining patterns contain rich phenotypic signatures that can be captured through radiomics and leveraged by machine learning models to predict relevant patient outcomes. This AutoML-based approach demonstrates high predictive performance in forecasting 6-month transplant eligibility, radiologic response, and survival metrics. These findings support the clinical integration of radiomics and AI to inform earlier, personalized decision-making following TACE and to optimize downstream management strategies in patients with HCC.