Establishment and validation of a CT-based clinical deep learning radiomics nomogram for predicting the response to transcatheter arterial chemoembolization in patients with hepatocellular carcinoma.

Background: Hepatocellular carcinoma (HCC) is a common malignant tumor of the liver worldwide, significantly impacting human health. Transcatheter arterial chemoembolization (TACE) is one of the key treatment modalities for HCC, making accurate prediction of the TACE response (TR) crucial for personalized treatment. This study aimed to develop and validate a clinical deep learning (DL) -based radiomics (Rad) nomogram (CDLRN) using pre-treatment contrast-enhanced computed tomography (CT) images and clinical data to predict TR in patients with HCC. Methods: Clinical and CT data were retrospectively collected from HCC patients who underwent TACE at the Department of Hepatobiliary Surgery of Shangluo Central Hospital from January 2019 to December 2023. A total of 144 patients met the study criteria and were randomly divided into training and validation cohorts at a ratio of 7:3, resulting in 102 and 42 patients, respectively. TR was assessed in according with the Modified Response Evaluation Criteria in Solid Tumors (mRECIST). Univariate and multivariate logistic regression (LR) analyses were conducted to identify independent risk factors associated with TR. Rad and DL features were extracted from preprocessed arterial-phase CT images. These features and image signatures were constructed through four steps: repeatability evaluation, univariate analysis, least absolute shrinkage and selection operator (LASSO) method, and multivariate LR analysis. A novel CDLRN model was then established to predict TR. Results: Among all models, the CDLRN significantly outperformed the others (P<0.05). Specifically, the area under the receiver operating characteristic curve (AUC) was 0.958 (95% confidence interval [CI]: 0.916-1.000) for the training cohort and 0.916 (95% CI: 0.835-0.997) for the validation cohort, both significantly higher than individual indicators. The CDLRN model demonstrated a specificity of 0.957, sensitivity of 0.839, and overall accuracy of 89.2% in the training cohort. In the validation cohort, the specificity was 1.0, sensitivity was 0.741, and accuracy was 83.3%. "The calibration curves demonstrated strong concordance between predicted and actual probabilities, underscoring the model's reliability. Furthermore, decision curve analysis highlighted the substantial clinical utility of the CDLRN model. Conclusion: The CDLRN showed high discrimination and accuracy in predicting the response to the first TACE in patients with HCC, demonstrating its potential to provide invaluable information for clinical decision-making.