Purpose: To identify radiomics subtypes that reflect tumor heterogeneity in bifocal hepatocellular carcinoma (bHCC) using an unsupervised machine learning approach. Additionally, to develop a preoperative model and a postoperative fusion model aimed at predicting recurrence-free survival (RFS) and overall survival (OS) in bHCC patients following hepatectomy. Methods: This retrospective study included 182 bHCC patients (91 in the training set, 91 in the test set). To capture the overall tumor characteristics, radiomics features were extracted from both lesions across six MR sequences and integrated using a two-lesion fusion approach to represent each patient as a single analytical entity. The similarity network fusion approach was utilized to construct a patient similarity matrix based on multi-sequence radiomic features, aiming to identify distinct subgroups that capture patterns of tumor imaging heterogeneity through spectral clustering. Multivariable Cox regression analysis was conducted to develop prognostic models for RFS and OS. The preoperative radiomics image heterogeneity (RIH) model and postoperative model including pathological features were built to predict prognosis of bHCC patients after hepatectomy. Results: Unsupervised clustering analysis based on multi-parametric radiomics revealed two subtypes correlated with distinct clinical outcomes, where high-radiomics image heterogeneity (high-RIH) was associated with poorer RFS (Log-rank p = 0.0059) and OS (Log-rank p = 0.0343). The independent predictors of shorter RFS included RIH cluster (HR, 1.782; 95% CI, 1.189– 2.670), pathological satellite nodule (HR, 1.946; 95% CI, 1.094– 3.460), MVI (HR, 1.714; 95% CI, 1.231– 2.386). The independent predictors of shorter OS included RIH cluster (HR, 2.008; 95% CI, 1.119– 3.605), radiological satellite nodule (HR, 1.982; 95% CI, 1.008– 3.901), MVI (HR, 4.350; 95% CI, 2.358– 8.028). Conclusion: This study identified two different radiomics subtypes in bHCC which could reveal the heterogeneity of bHCC and predict clinical outcomes in post-hepatectomy bHCC patients. The diagram illustrates a radiomics workflow. It begins with imaging acquisition, showing a group of people and an MRI scan. This leads to tumor segmentation, depicted with a segmented tumor on an MRI image. Next is radiomics features extraction, including shape, histogram and texture features. This information is used in similarity network fusion, represented by a network diagram and a heatmap. The process results in risk stratification into high and low risk, labeled as High-RIH and Low-RIH. Each risk group is associated with icons of people, a checklist, a microscope and test tubes. Two survival probability graphs are shown: one for high risk (Cluster 1) and one for low risk (Cluster 2), both plotting survival probability against follow-up time in months, with a log-rank p-value of 0.001 for high risk.Radiomics: imaging, feature extraction, fusion, risk stratification, survival graphs. Keywords: hepatocellular carcinoma, magnetic resonance imaging, unsupervised machine learning, prognosis
Jia et al. (Fri,) studied this question.