MRI-based radiomics with automated segmentation and manual refinement for predicting treatment response in unresectable hepatocellular carcinoma: a multicenter study.
Authors
Affiliations (6)
Affiliations (6)
- First Affiliated Hospital of GuangXi Medical University, Nanning, China.
- Qinzhou Second People's Hospital, Qinzhou, China.
- Qinzhou First People's Hospital, Qinzhou, China.
- The First Affiliated Hospital of Guilin Medical University, Guilin, China.
- First Affiliated Hospital of GuangXi Medical University, Nanning, China. [email protected].
- Key Laboratory of Early Prevention and Treatment for Regional High Frequency Tumor, Nanning, China. [email protected].
Abstract
To compare the effect of different segmentation strategies on the performance of multiparametric MRI-based radiomics models for predicting response to combined therapy in unresectable hepatocellular carcinoma (HCC). This multicenter retrospective study included 334 patients with HCC from five tertiary hospitals between January 2021 and January 2025. Clinical data and pre-treatment MRI were collected for all eligible patients. Tumor regions of interest (ROIs) were automatically segmented using a DeepLabV3-ResNet50 deep convolutional neural network. This segmentation model was applied separately to each MRI sequence. A radiologist reviewed the machine-generated ROIs and, when necessary, refined them. Based on three different segmentation strategies, three radiomics models were constructed: a fully automated model (Auto), a manual segmentation-based model (Manual), and a hybrid model based on automated segmentation followed by manual refinement (Auto + Refine). Patients from Hospitals 1-3 were randomly assigned to training (70%) and test (30%) cohorts, and an external validation cohort was formed from Hospitals 4 and 5. The mean Dice similarity coefficient (DSC) of the T2-FS automated segmentation model was 0.926 (95% CI, 0.899-0.951), that of the DWI automated segmentation model was 0.920 (95% CI, 0.886-0.944), and that of the T1-DCE automated segmentation model was 0.910 (95% CI, 0.893-0.928). The fully automated radiomics model achieved AUCs of 0.804 (95% CI, 0.694-0.913) in the test set and 0.735 (95% CI, 0.573-0.896) in the external validation cohort. The manual model achieved AUCs of 0.868 (95% CI, 0.756-0.979) and 0.876 (95% CI, 0.767-0.985), respectively. The hybrid model (Auto + Refine) achieved AUCs of 0.885 (95% CI, 0.732-0.954) and 0.893 (95% CI, 0.789-0.985), respectively. Hybrid (Auto + Refine) segmentation outperforms purely manual and fully automatic approaches in predicting treatment response in unresectable HCC. In settings where both manual annotation burden and accuracy are taken into consideration, automatic segmentation combined with manual refinement could serve as a preferred workflow.