Creation of a Prediction Model of Local Tumor Recurrence After a Successful Conventional Transcatheter Arterial Chemoembolization Using Cone-Beam Computed Tomography Based–Radiomics

To create and evaluate prediction models of local tumor recurrence after successful conventional transcatheter arterial chemoembolization (c-TACE) via radiomics analysis of lipiodol deposition using cone-beam computed tomography (CBCT) images obtained at the completion of TACE.

A total of 103 hepatocellular carcinoma nodules in 71 patients, who achieved a complete response (CR) based on the modified Response Evaluation Criteria in Solid Tumors 1 month after TACE, were categorized into two groups: prolonged CR and recurrence groups. Three types of areas were segmented on CBCT: whole segment (WS), tumor segment (TS), and peritumor segment (PS). From each segment, 105 radiomic features were extracted. The nodules were randomly divided into training and test datasets at a ratio of 7:3. Following feature reduction for each segment, three models (clinical, radiomics, and clinical–radiomics models) were developed to predict recurrence based on logistic regression.

The clinical–radiomics model of WS showed the best performance, with the area under the curve values of 0.853 (95% confidence interval: 0.765–0.941) in training and 0.752 (0.580–0.924) in test dataset. In the analysis of radiomic feature importance of all models, among all radiomic features, glcm_MaximumProbability, shape_MeshVolume and shape_MajorAxisLength had negative coefficients. In contrast, shape_SurfaceVolumeRatio, shape_Elongation, glszm_SizeZoneNonUniformityNormalized, and gldm_GrayLevelNonUniformity had positive coefficients.

In this study, a machine-learning model based on cone-beam CT images obtained at the completion of c-TACE was able to predict local tumor recurrence after successful c-TACE. Nonuniform lipiodol deposition and irregular shapes may increase the likelihood of recurrence.