This study proposes a framework integrating medical foundation segmentation models and radiomics for liver lesion segmentation and classification in multi-modal magnetic resonance imaging (MRI). MedSAM2 and VISTA3D are complementarily fine-tuned and combined through spatial gated fusion to generate lesion masks. Deep semantic features extracted by M3D are integrated with handcrafted radiomics features through within-modality alignment, cross-modal interaction, and sparse selection. Negative sampling and an extreme gradient boosting (XGBoost) classifier are then used for lesion classification. On the validation set of the public LLD-MMRI2023 dataset, the framework achieved a Dice coefficient of 0.90, an F1 score of 0.837, and a Cohen’s Kappa of 0.792. The latter two metrics represented relative improvements of 8.27% and 7.57% over the best known public results. Ablation experiments demonstrated the contributions of joint fine-tuning, spatial priors, and multi-modal feature fusion. The proposed framework combines foundation segmentation models, deep semantic features, and radiomics to support liver lesion segmentation and classification in multi-modal MRI.
