Longitudinal CE-MRI-based Siamese network with machine learning to predict tumor response in HCC after DEB-TACE

Abstract

Background: Accurate prediction of tumor response after drug-eluting beads transarterial chemoembolization (DEB-TACE) remains challenging in hepatocellular carcinoma (HCC), given tumor heterogeneity and dynamic changes over time. Existing prediction models based on single timepoint imaging do not capture dynamic treatment-induced changes. This study aims to develop and validate a predictive model that integrates deep learning and machine learning algorithms on longitudinal contrast-enhanced MRI (CE-MRI) to predict treatment response in HCC patients undergoing DEB-TACE.

Methods: This retrospective study included 202 HCC patients treated with DEB-TACE from 2004 to 2023, divided into a training cohort (n = 141) and validation cohort (n = 61). Radiomics and deep learning features were extracted from standardized longitudinal CE-MRI to capture dynamic tumor changes. Feature selection involved correlation analysis, minimum redundancy maximum relevance, and least absolute shrinkage and selection operator regression. The patients were categorized into two groups: the objective response group (n = 123, 60.9%; complete response = 35, 28.5%; partial response = 88, 71.5%) and the non-response group (n = 79, 39.1%; stable disease = 62, 78.5%; progressive disease = 17, 21.5%). Predictive models were constructed using radiomics, deep learning, and integrated features. The area under the receiver operating characteristic curve (AUC) was used to evaluate the performance of the models.

Results: We retrospectively evaluated 202 patients (62.67 ± 9.25 years old) with HCC treated after DEB-TACE. A total of 7,182 radiomics features and 4,096 deep learning features were extracted from the longitudinal CE-MRI images. The integrated model was developed using 13 quantitative radiomics features and 4 deep learning features and demonstrated acceptable and robust performance with an receiver operating characteristic curve (AUC) of 0.941 (95%CI: 0.893–0.989) in the training cohort, and AUC of 0.925 (95%CI: 0.850–0.998) with accuracy of 86.9%, sensitivity of 83.7%, as well as specificity of 94.4% in the validation set.

Conclusions: This study presents a predictive model based on longitudinal CE-MRI data to estimate tumor response to DEB-TACE in HCC patients. By capturing tumor dynamics and integrating radiomics features with deep learning features, the model has the potential to guide individualized treatment strategies and inform clinical decision-making regarding patient management.

Supplementary Information: The online version contains supplementary material available at 10.1186/s40644-025-00926-5.

Keywords: Deep learning; HCC; Machine learning; Siamese network; Tumor response.

Fig. 1

The study design and workflow of longitudinal CE-MRI-based Siamese network with machine learning to predict tumor response in HCC after DEB-TACE. SDF Siamese delta features; Lasso Least absolute shrinkage and selection operator; MSE Minimum mean squared error; DRF Delta radiomic features; SVM, support vector machine; KNN, nearest neighbors; XGBoost, extreme gradient boosting; LightGBM, light gradient boosting machine;

Fig. 2

Radiomic feature selection using the Maximum relevance minimum redundancy (MRMR) and the least absolute shrinkage and selection operator (LASSO). (a) A clustering heatmap of Spearman rank correlation selected features deriving from longitudinal CE-MRI images. The resulting heatmap is visualized using a color bar of the z-score, a higher z-score with a more red display and a lower z-score with a more blue display. (b) Selection of the tuning parameter (λ) in the LASSO model using five-fold cross-validation via minimum mean squared error (MSE). Dotted vertical lines drawn at the optimal values of 0.0126 were selected; (c) The histogram exhibits radiomics features contributed to the constructed radiomics model. The y-axis represents radiomics features, with their coefficients in the multivariate logistic regression analysis plotted on the x-axis. (d) LASSO coefficient profiles of the radiomic features. The dotted vertical line was plotted at the optimal λ, resulting in 17 features with nonzero coefficients

Fig. 3

The receiver operating characteristic (ROC), confusion matrix of the different models in the training and validation cohorts (a-i). The ROC curves of the radiomics model, deep learning model, and integrated model in the training (a, e, i) and validation (c, g, k) cohorts. The confusion matrix of the radiomics model, deep learning model, and integrated model in the training (b, f, j) and validation (d, h, l) cohorts

Fig. 4

Integrated delta SVM model explainability. (a) SHAP beeswarm plots of the model. The plot illustrated the feature relevance and combined feature attributions to the model’s predictive performance. CE-MRI images show the tumor in the arterial phase (AP), portal venous phase (PVP) and delayed phase (DP) before and after DEB-TACE from a 69-year-old man with a 119 mm lesion in the left lobe of the liver (b) and a 72-year-old man with a 35 mm lesion in the right lobe of the liver (c). The SHAP waterfall plot summarizes the contribution of individual imaging features to the model’s prediction for this patient, with the x-axis representing the model output and features ranked by their impact. Features with negative SHAP values (blue) decrease the prediction score, while features with positive SHAP values (red) increase it. The SHAP force plot further illustrates the cumulative effect of all features, starting from the baseline prediction (E[f(x)] = 0.65) to the final model output