Radiomics Integrated with Deep Learning Based on Contrast-Enhanced MRI for Predicting Short-Term Efficacy of...

We aimed to develop and validate a predictive model combining radiomics, deep learning, and clinical features for the preoperative prediction of the short-term efficacy of initial drug-eluting bead transarterial chemoembolization (DEB-TACE) treatment in patients with hepatocellular carcinoma (HCC). This retrospective cohort study included 113 internal and 32 external patients from three centers with intermediate and advanced HCC were included. A total of 79 internal cases were included in the training group, 34 in the internal testing group, and 32 in the external testing group. We extracted 1496 radiomics features per sequence based on three sequences: arterial phase (AP), diffusion-weighted imaging (DWI), and T2-weighted imaging (T2WI) based on contrast-enhanced magnetic resonance imaging. Deep learning features were extracted using a pretrained 3D ResNet-18 architecture to analyze AP images. Specifically, the activation of 512 neurons from the final fully connected layer of the network was employed as a deep learning feature. Minimum redundancy maximum correlation and least absolute shrinkage and selection operator (LASSO) regression were employed for feature selection and model construction. In this study, the following six models were developed: three models based on individual sequences (AP, DWI, and T2WI); a deep learning model using AP images; a two-sequence model combining AP and DWI; and a comprehensive model integrating radiomics, deep learning, and clinical features. A multifactor logistic regression was used to develop a clinical imaging model based on clinical factors, and a comprehensive model based on radiomics, deep learning, and clinical features. Model performance was assessed using the area under the curve (AUC) and calibration curves, while decision curve analysis (DCA) was used to evaluate the clinical value of the model. A total of 34 radiomics features were selected using LASSO regression. In the training group, the best predictive performance was achieved by using the single AP model (training set AUC = 0.900; testing set AUC = 0.750). In the testing group, the AP- and DWI-based combined radiomics, deep learning, and clinical model had the best predictive performance (training set AUC = 0.864, internal testing set AUC = 0.822, and external testing set AUC = 0.804). Multifactorial logistic regression analysis showed that microsphere type (p = 0.042), deep learning, and AP- and DWI-based RAD-scores (p = 0.01) were associated with short-term efficacy. Subsequently, a nomogram was constructed to provide a more visual representation. The comprehensive model based on radiomics, deep learning, and clinical features can effectively predict short-term efficacy of initial DEB-TACE in patients with HCC.