Dynamic Vascular Spatiotemporal Heterogeneity on Multiphase CT for Preoperative Prediction of Microvascular Invasion in Hepatocellular Carcinoma.
Authors
Affiliations (6)
Affiliations (6)
- Department of Radiology, The First Affiliated Hospital, Jiangxi Medical College, Nanchang University, Nanchang, 330006, People's Republic of China.
- Clinical Research Center for Medical Imaging of Jiangxi Province, Nanchang, 330006, People's Republic of China.
- Department of Radiology, Jiangxi Cancer Hospital, Nanchang, 330006, People's Republic of China.
- Department of Radiology, The Second Affiliated Hospital of Nanchang University, Nanchang, 330006, People's Republic of China.
- Department of Pathology, The First Affiliated Hospital of Nanchang University, Nanchang, 330006, People's Republic of China.
- Jiangxi Provincial Key Laboratory for Precision Pathology and Intelligent Diagnosis, Department of Pathology and Institute of Molecular Pathology, The First Affiliated Hospital, Jiangxi Medical College, Nanchang University, Nanchang, 330006, People's Republic of China.
Abstract
Microvascular invasion (MVI) is an important determinant of postoperative recurrence in hepatocellular carcinoma (HCC) but is usually confirmed only after surgery. This study aimed to develop and externally validate a multiphase CT framework integrating dynamic vascular trajectory, spatial intratumoral heterogeneity, and deep imaging features for preoperative MVI prediction and exploratory prognostic stratification. This retrospective multicenter study included 405 patients with pathologically confirmed HCC: 198 in the development cohort and 207 in two independent external-validation cohorts. Clinical variables, static multiphase radiomics, Delta vascular trajectory features, three-dimensional intratumoral heterogeneity (3D-ITH)/habitat features, and Vision Transformer (ViT) features were extracted. All preprocessing, model fitting, and threshold selection were confined to the development cohort. A capacity-constrained multimodal Transformer was compared with clinical, single-modality, simple concatenation, and tree-based models. The cohort included 148 MVI-positive and 257 MVI-negative patients according to the locked blinded multireviewer reference standard. In pooled external validation, Transformer fusion achieved an area under the curve of 0.941 (95% CI, 0.906-0.969), sensitivity of 0.951, specificity of 0.754, accuracy of 0.831, negative predictive value of 0.960, and Brier score of 0.113. Corresponding AUCs were 0.957 and 0.934 in the two external cohorts. Delta-only and clinical models achieved pooled external AUCs of 0.793 and 0.675, respectively. High fusion-risk patients had poorer composite RFS/PFS than low-risk patients (hazard ratio, 2.05; 95% CI, 1.36-3.08; P < 0.001). The proposed dynamic vascular heterogeneity-centered framework demonstrated high external discrimination for preoperative MVI prediction in HCC and provided exploratory stratification of the available composite RFS/PFS endpoint. Prospective validation and recalibration are required before clinical implementation.