BioFact-MoE: AI Framework for Liver Cancer Prognosis
BioFact-MoE is a framework that utilizes a biologically factorized Mixture of Experts for prognostic modeling related to vision and language in hepatocellular carcinoma (HCC). This model distinctly separates liver and tumor factors through biologically supervised experts integrated within a residual MoE survival architecture, enhancing both accuracy and biological interpretability compared to current VLMs that rely on a single, intertwined latent representation. The model underwent pretraining with 4,582 pairs of 3D MRI images and reports and was evaluated using a group of 588 HCC patients.
Key facts
- BioFact-MoE is a biologically factorized Mixture of Experts framework.
- It is designed for vision-language prognostic modeling in hepatocellular carcinoma (HCC).
- It decomposes liver and tumor factors via biologically supervised experts.
- It uses a residual MoE survival architecture.
- Existing VLMs learn a single entangled latent representation.
- BioFact-MoE improves accuracy and biological interpretability.
- Pretrained on 4,582 3D MRI image-report pairs.
- Tested on a cohort of 588 HCC patients.
Entities
—