Machine Learning using MR İmaging Radiomics can Predict the Response of Large Hepatocellular Carcinoma to Transarterial Radioembolization.

BACKGROUND

Large tumor size is associated with poor outcomes in patients with hepatocellular carcinoma (HCC). Although some studies have evaluated the treatment response of HCC to transarterial radioembolization (TARE), none of them used radiomics features with machine learning (ML) models in large tumors.

AIM: To assess the performance of ML models using radiomics to predict the treatment response of TARE in large HCC lesions.

METHODS

This study included 49 patients with a large (>5 cm) HCC who underwent TARE. Treatment response was determined according to modified response evaluation criteria in solid tumors (mRECIST) criteria from the 3-month follow-up MR examinations. Complete or partial response was categorized as the responder group, while stable or progressive disease was classified as the non-responder group. Segmentation was performed on axial T2-weighted (T2W) and contrast-enhanced (CE) T1-weighted images. Classification learning models were used to create prediction models for TARE response.

RESULTS

Forty-nine patients (9 female, 40 male; mean age 63.58 ± 8.77) were included. None of the clinical, laboratory, and radiologic characteristics except the neutrophil counts showed statistical significance. Radiomics models obtained from CE-T1 and T2W images demonstrated an accuracy rate of 79.6%, while the area under the curve (AUC) rates were 0.92 and 0.77, respectively. The clinical model showed an accuracy rate of 77.6% and an AUC of 0.65. No statistically significant difference was found among all the models (P = 0.092).

CONCLUSION

ML-based models constructed with radiomics features obtained from MR images before the TARE procedure might predict response in large HCC lesions.