Integrated multi-task learning framework for hepatocellular carcinoma segmentation and histological grading using fused multi-phase MRI.
Authors
Affiliations (3)
Affiliations (3)
- The Second Affiliated Hospital, Department of Ultrasonography, Hengyang Medical School, University of South China, Hengyang, China.
- The Second Affiliated Hospital, Department of electrocardiogram, Hengyang Medical School, University of South China, Hengyang, China.
- The Second Affiliated Hospital, Department of Ultrasonography, Hengyang Medical School, University of South China, Hengyang, China. [email protected].
Abstract
This study aims to develop and validate an integrated multi-task framework for hepatocellular carcinoma analysis by combining deep learning-based segmentation with radiomics-based histological grading using fused multi-phase MRI. In this retrospective study, MRI data from 1673 patients with histopathologically confirmed hepatocellular carcinoma (875 high-grade, 798 low-grade) were analyzed. Arterial-phase and portal-venous-phase T1-weighted images were acquired using a standardized, bolus-tracking protocol to ensure consistent contrast timing. Six segmentation models, Vision Transformer, nnU-Net, U-Net, DeepLabV3+, Swin Transformer, and SegNet, were trained on arterial, portal-venous, and fused (wavelet-based) MRI data. Radiomic features (n = 215) were extracted from segmented tumor volumes and pre-filtered to remove multicollinearity. Feature refinement was performed using Lasso, Recursive Feature Elimination (RFE), and ANOVA. Tumor grade classification was conducted using TabTransformer, TabNet, XGBoost, and CatBoost. Five-fold cross-validation and an independent test set were used for robust evaluation. Standardized preprocessing, including intensity normalization, bias field correction, and inter-phase registration, ensured consistent image quality and analytical reproducibility. The proposed framework achieved high segmentation accuracy with DSC scores above 0.92 across fused MRI images. Classification performance was exceptional, with training accuracy reaching 93.2% and testing accuracy 92.5%, while AUC values exceeded 96% in the fused modality. Comparative analyses revealed that the Transformer-RFE-Fused model outperformed alternative architectures, demonstrating superior generalization and robust feature learning. In addition, SHAP analysis confirmed the high contribution of key radiomic features, and t-SNE visualizations illustrated clear separation between low-grade and high-grade tumors. These results validate the efficacy of our multi-task learning approach in enhancing HCC tumor segmentation and grading. Our evaluation underscores the clinical potential of our integrated framework for accurate, reproducible, and interpretable HCC diagnosis. Our integrated multi-task learning framework markedly improves HCC tumor segmentation and grading. Transformer-RFE-Fused (Wavelet) MRI offers superior accuracy and robustness, efficiently paving the way for enhanced clinical decision-making.