Development and Validation of a Radiomics Nomogram Based on Magnetic Resonance Imaging and Clinicoradiological Factors to Predict HCC TACE Refractoriness.

Purpose: This study constructs a predictive model for hepatocellular carcinoma (HCC) transarterial chemoembolization (TACE) refractoriness using a machine learning (ML) algorithm and verifies the predictive performance of different algorithms. Patients and Methods: Clinical and magnetic resonance imaging (MRI) data of 131 patients (48 with TACE refractoriness) who underwent repeated TACE treatment for HCC were retrospectively collected. The training and validation cohorts comprised 104 and 27 cases, respectively, following an 8:2 ratio. Clinical imaging characteristics related to TACE refractoriness were identified through logistic regression analysis. HCC lesions on arterial phase, portal phase, delayed phase, and T2-weighted fat suppression MRI images before the first TACE were manually delineated as regions of interest. Dimension reduction was conducted using variance threshold, univariate selection, and least absolute shrinkage and selection operator methods. Relevant indices of TACE refractoriness were selected. ML algorithms, including a support vector machine, random forest, logistic regression and adaptive boosting, were used to construct the radiomics, clinical prediction, and combined models. The predictive performance of these models was evaluated using receiver operating characteristic curves. The optimal model was presented as a nomogram and verified through calibration and decision curve analyses. Results: In evaluating radiomics models for predicting TACE refractoriness in HCC, the LR-developed portal venous phase (VP) model achieved optimal single-sequence performance (training AUC: 0.896, 95% CI: 0.843-0.941; validation: 0.853, 0.727-0.965). Multisequence models significantly surpassed single-sequence counterparts, with the T2WI-FS+AP+VP+DP multisequence LR model demonstrating peak efficacy (training: 0.905, 0.853-0.949; validation: 0.876, 0.773-0.976). The integrated clinical-radiomics model demonstrated robust predictive performance, achieving a training cohort AUC of 0.955 (95% CI: 0.918-0.984) with 0.885 accuracy, 0.921 sensitivity, and 0.864 specificity, and maintained strong validation performance (AUC=0.941, 95% CI: 0.880-0.991). Conclusion: Multisequence clinical-radiomics model accurately predicts TACE refractoriness in hepatocellular carcinoma.