MeWM - An AI-powered medical world model that accurately simulates tumor evolution.

What is MeWM?

MeWM (Medical World Model) is an innovative medical model proposed by institutions such as the Hong Kong University of Science and Technology (Guangzhou) to assist clinical decision-making by simulating disease dynamics. It consists of a strategy model, a dynamic model, and an inverse dynamic model. The strategy model, based on a vision-language model, generates potential treatment plans according to the patient's state; the dynamic model simulates the progression or regression of the disease under different treatment conditions; and the inverse dynamic model assesses the survival risk of the simulation results and optimizes the treatment plan. MeWM uses medical images as input, constructs the initial disease state through a perception module, predicts the future disease state through the dynamic model, evaluates the treatment effect through the inverse dynamic model, and generates the optimal treatment plan through the strategy model, forming a closed-loop feedback to optimize the intervention path.

MeWM's main functions

• Tumor evolution simulationMeWM can use 3D diffusion models to simulate the evolution of tumor morphology under different treatment options, generating realistic postoperative tumor images to help doctors intuitively assess treatment effectiveness before surgery.
• Survival risk assessmentThrough survival analysis models, MeWM can predict the prognostic survival risk of patients under different treatment regimens and deduce the optimal treatment plan based on inverse dynamics. The evaluation results significantly outperform traditional multimodal large models in terms of accuracy.
• Clinical decision optimizationMeWM constructs an automated, visualized optimization loop from "treatment plan generation - simulation and simulation - survival assessment" to support interventional cancer treatment planning. It can significantly improve the clinical decision-making ability of interventional physicians; for example, in the selection of treatment plans for transarterial chemoembolization (TACE) in liver cancer, it improved the F1 score by 13%.

MeWM's technical principles

• Technical ArchitectureMeWM consists of the following three core components:
  • Policy ModelBased on vision-language models (such as GPT-4o), candidate treatment combinations that meet clinical rules are generated from CT images and treatment goals, covering a variety of therapeutic drugs and embolization materials.
  • Dynamics ModelUsing a 3D conditional diffusion model, the morphological changes of tumors after surgery are simulated according to different treatment plans, generating realistic postoperative tumor images.
  • Inverse Dynamics ModelThe survival risk of simulated postoperative tumor images is assessed, the risk value is calculated using a heuristic function, and the treatment plan is dynamically optimized. Finally, the plan with the lowest risk is selected as the recommendation.
• Workflow
  • Input and perceptionUsing medical images as input, the initial condition status is generated through the perception module.
  • Simulation and EvaluationThe dynamic model simulates the morphological changes of the tumor after surgery based on different treatment plans, and the inverse dynamic model assesses the survival risk of the simulation results.
  • Optimization and FeedbackBased on the evaluation results, the strategy model generates new treatment plans, forming a closed-loop feedback loop to continuously optimize the intervention path.
• Innovation
  • Generative modelsMeWM innovatively uses "treatment behavior" as a conditional input to generate a model, simulating tumor evolution under different treatment plans.
  • Comparative learningBy combining contrastive learning (CCL), the realism and treatment sensitivity of the generated results are enhanced.
  • Clinical decision supportIn the selection of treatment options for transarterial chemoembolization (TACE) in liver cancer, MeWM significantly improved the decision-making accuracy of interventional physicians, increasing the F1 score by 13%.

MeWM project address

• Project official websitehttps://yijun-yang.github.io/MeWM/
• Github repositoryhttps://github.com/scott-yjyang/MeWM
• arXiv technical paper: https://arxiv.org/pdf/2506.02327

Application scenarios of MeWM

• Preoperative planning and treatment plan optimizationMeWM provides doctors with intelligent assistance for preoperative planning. By generating postoperative tumor images and assessing survival risk, MeWM helps doctors optimize treatment plans and select the most suitable combination of drugs and embolization materials for each patient.
• As a "second radiologist"MeWM holds promise for integration into clinical practice as a "second radiologist." It can perform survival analysis and scoring based on generated postoperative CT scans, forming a complete closed-loop mechanism of "treatment simulation—effect evaluation—treatment plan selection," assisting physicians in more effectively predicting treatment outcomes before surgery.
• Personalized treatment supportMeWM supports personalized, data-driven treatment decisions. It can generate multiple treatment options based on a patient's preoperative imaging and treatment goals, and select the optimal option through simulation and evaluation, providing strong support for precision medicine.