Enhancing TACE Response Prediction for Hepatocellular Carcinoma: A Deep Learning Framework Based on Pre-Procedural CT and Modified AlexNet

Early response assessment to transcatheter arterial chemoembolization (TACE) is paramount for Hepatocellular carcinoma (HCC) patients since the tumor response to TACE informs the decision about further therapeutic planning, such as treatment or switching to alternative therapeutic modalities. Due to the complex heterogeneity of tumors, reliance on size-based assessment for TACE response prediction, such as using imaging criteria like RECIST, has limitations in capturing changes in tumor vascularity, necrosis, and viability. Considering these, this study aimed to develop and evaluate an intelligent system for the initial response prediction to TACE based on the convolutional neural network (CNN). 69 HCC patients' pre-procedural CT scans and clinical data were used in the study. The TACE responses were categorized into two groups based on clinical data: refractory and susceptibility. Several CNN models were trained using transfer learning due to the small dataset. Among them, modified AlexNet demonstrated state-of-art performance in forecasting TACE response with accuracies of 100% for training, 100% for validation and 99.27% for testing. The AUC score was 0.992, with an average recall, precision, and F1-score of 99.49%, 98.99%, and 99.23% on the test set. The experimental results of this study highlight the value of using modified pre-trained models in predicting the TACE responses with the highest accuracy.