Predicting the Initial Treatment Response to Transarterial Chemoembolization in Intermediate-Stage Hepatocellular Carcinoma by the Integration of Radiomics and Deep Learning

Abstract

Objectives: We aimed to develop radiology-based models for the preoperative prediction of the initial treatment response to transarterial chemoembolization (TACE) in patients with hepatocellular carcinoma (HCC) since the integration of radiomics and deep learning (DL) has not been reported for TACE.

Methods: Three hundred and ten intermediate-stage HCC patients who underwent TACE were recruited from three independent medical centers. Based on computed tomography (CT) images, recursive feature elimination (RFE) was used to select the most useful radiomics features. Five radiomics conventional machine learning (cML) models and a DL model were used for training and validation. Mutual correlations between each model were analyzed. The accuracies of integrating clinical variables, cML, and DL models were then evaluated.

Results: Good predictive accuracies were showed across the two cohorts in the five cML models, especially the random forest algorithm (AUC = 0.967 and 0.964, respectively). DL showed high accuracies in the training and validation cohorts (AUC = 0.981 and 0.972, respectively). Significant mutual correlations were revealed between tumor size and the five cML models and DL model (each P < 0.001). The highest accuracies were achieved by integrating DL and the random forest algorithm in the training and validation cohorts (AUC = 0.995 and 0.994, respectively).

Conclusion: The radiomics cML models and DL model showed notable accuracy for predicting the initial response to TACE treatment. Moreover, the integrated model could serve as a novel and accurate method for prediction in intermediate-stage HCC.

Keywords: TACE; deep learning; hepatocellular carcinoma; machine learning; treatment response.

Figure 1

Flowchart for the integration of the machine learning and deep learning models. (i) The radiologists manually segmented the 3D ROIs of HCC using ITK-SNAP. Thereafter, 1167 features were extracted from the hepatic arterial CT images based on the “pyradiomics” package of python. Redundant radiomics features were then eliminated by ICCs. Using RFE and 5-fold cross-validation, the final features were selected and the five radiomics models were built using different machine learning algorithms. (ii) A CT image and mask of ROI manually segmented from the largest tumor layer was resized and output as 224 × 224 ×3 from each patient in the two cohorts. Augmentation techniques were used to enlarge the training CT image dataset, and “new” big data were generated. The deep learning framework included two convolutions, two max-poolings, and one dense layer. The final output layer was a softmax classifier. The optimizer was Adam, with a learning rate of 0.001 and batch size of 16. All the layers were standardized and the L2 regularization was set to 0.000001. The activation function of RELU was set as alpha = 0.1. (iii Each of the five machine learning models (linear, logistic, GBM, SVM, and RF) was integrated with the deep learning model to predict the initial treatment response to TACE. CT, computed tomography; GBM, gradient boosting machine; HCC, hepatocellular carcinoma; ICC, intraclass correlation coefficient; RELU, reconstructed linear units; RF, random forest; RFE, recursive feature elimination; ROI, region of interest; SVM, support vector machine; TACE, transarterial chemoembolization.

Figure 2

Associations between the clinical factors and initial treatment response to TACE. (A) The CT images from a patient acquiring CR after one course of TACE treatment were presented. Clinical factors predicting response (CR + PR) in the training (B) and validation cohorts (C). CR, complete response; PR, partial response; TACE, transarterial chemoembolization.

Figure 3

Five radiomics cML models and the DL model could precisely predict initial treatment response to TACE. ROC curves showing the predictive performance of the five cML models for estimating treatment response in the training (A) and validation cohorts (B). ROC curves showing the predictive performance of the DL model for predicting treatment response in the training (C) and validation (D) cohorts. cML, conventional machine learning; DL, deep learning; ROC, receiver operating characteristic; TACE, transarterial chemoembolization.

Figure 4

Correlation between tumor size, cML, and DL. Evaluating the mutual correlation between tumor size, the five radiomics cML models, and the DL model via correlation heatmaps in the training (A) and validation (B) cohorts. cML, conventional machine learning; DL, deep learning.

Figure 5

Integrating tumor size, cML models, and DL model to predict initial treatment response. Predictive performances of the ensemble model, including tumor size and the cML and DL models, are shown as ROC curves for the training (A) and validation (B) cohorts. cML, conventional machine learning; DL, deep learning; ROC, receiver operating characteristic.