Transformer-Based Habitat Fusion Model Using DCE-MRI Predicts Microvascular Invasion and Recurrence in Hepatocellular Carcinoma.
Authors
Affiliations (5)
Affiliations (5)
- Department of Magnetic Resonance, The Second Hospital & Clinical Medical School, Lanzhou University, Lanzhou, Gansu, People's Republic of China.
- Gansu Province Clinical Research Center for Functional and Molecular Imaging, Lanzhou, Gansu, People's Republic of China.
- Gansu Medical MRI Equipment Application Industry Technology Center, Lanzhou, Gansu, People's Republic of China.
- Department of Radiology, Huaxi MR Research Center (HMRRC), Institute of Radiology and Medical Imaging, Psychoradiology Key Laboratory of Sichuan Province, West China Hospital of Sichuan University, Chengdu, Sichuan, People's Republic of China.
- Department of Nuclear Medicine, The Second Hospital & Clinical Medical School, Lanzhou University, Lanzhou, Gansu, People's Republic of China.
Abstract
Microvascular invasion (MVI) is the primary risk factor driving recurrence and metastasis of hepatocellular carcinoma (HCC) after surgical treatment. We developed a fusion model integrating dynamic contrast-enhanced MRI (DCE-MRI)-derived habitat radiomics and conventional radiomics with clinical features to preoperatively predict MVI status and recurrence-free survival (RFS). This study included 193 HCC patients from two centers. Logistic regression analysis was used to identify the independent clinical predictors of MVI. Additionally, habitat radiomics and conventional radiomics were derived from the DCE-MRI data. A transformer deep-learning model was used to integrate multimodal habitat features. Subsequently, a habitat model, a radiomics model and a clinical model were constructed. These models were then integrated into different fusion models. The performance of the single and fusion models was evaluated using different metrics, and the RFS was assessed using Kaplan-Meier analysis. Multivariate analysis identified pseudocapsule integrity and tumor diameter as independent clinical predictors of MVI. Transformer outperformed traditional machine learning methods in fusing multimodal habitat features. A fusion model integrating habitat radiomics, conventional radiomics and clinical predictors demonstrated the best performance (training data: area under the curve [AUC] = 0.950, 95% confidence interval [CI]: 0.918-0.981; testing data: AUC = 0.923, 95% CI: 0.858-0.988) for predicting MVI, with robust calibration and satisfactory clinical utility. Kaplan-Meier analysis confirmed significant RFS stratification for both MVI-positive (<i>p</i> = 0.038) and high-risk nomogram groups (<i>p</i> = 0.033). A fusion model integrating DCE-MRI-based habitat radiomics and conventional radiomics with clinical features provides excellent prediction of MVI (AUC = 0.923) and RFS in patients with HCC. This model may provide important support for optimized surgical planning and personalized therapy.