Abstract
OBJECTIVE: This study aims to develop a multimodal model that integrates clinical features and preoperative imaging characteristics to predict complete response (CR) in hepatocellular carcinoma (HCC) patients following ablation therapy, and to assess the therapeutic efficacy of ablation therapy. METHODS: From October 2017 to June 2024, we collected clinical data and CT enhanced images from 108 HCC patients within one month before ablation therapy. The most important features were selected and dimensionality was reduced using the Mann-Whitney U test, Principal Component Analysis (PCA), and Least Absolute Shrinkage and Selection Operator (LASSO) regression. A multimodal feature set was constructed by integrating clinical characteristics. The dataset is randomly divided into training and validation sets at a 7:3 ratio. Using 10-fold cross-validation, we employ eight radiomic and deep learning algorithms (such as logistic regression, random forest, and MLP) to train models for predicting the CR of HCC. The performance of the optimal model is then comprehensively evaluated. RESULTS: A total of 14 clinical datasets were collected, resulting in the extraction of 3,192 radiomic features and 256 deep learning features. Among the clinical datasets, a significant difference was found in the rate of HBV positivity (p<0.05). From the radiomic features, 26 key features were selected and dimensionally reduced, while 5 key features were selected and reduced from the deep learning features. The multimodal feature set (Radiomics + Clinical + Deeplearning) achieved the best AUC results across MLP and several other machine learning models. Notably, the MLP model delivered the highest overall performance, with an AUC of 0.933 in the test set. The accuracy, specificity, and positive predictive value of the MLP model were 0.79, 0.99, and 0.89, respectively. CONCLUSION: This pilot study established a multimodal model combining radiomics and deep learning to predict the response to ablation therapy in HCC. The model demonstrated robust performance, providing a reliable tool for personalized efficacy assessment and individualized treatment strategies.