NA-MIC Project Weeks

Back to Projects List

AI driven interface for SlicerTMS

Key Investigators

• SangHyuk Kim (BWH & UMass Boston, USA)
• Puxun Tu (BWH & SJTU, USA)
• Steve Pieper (Isomics, USA)
• Lipeng Ning (BWH & HMS, USA)

Project Description

SlicerTMS is a 3DSlicer module for patient-specific transcranial stimulation. It integrates several functions, including neuronavigation, electric field modeling, real-time EEG streaming and recording, and TMS control. These functions involve a complex user interface, and some tasks, such as neuronavigation registration, may need more than one user. To simplify the interface and improve the user experience, we will develop a new version leveraging LLM models and Slicer AI Agent tools. Specifically, we will eliminate LLM hallucinations at the infrastructure level by executing medical software APIs through human-verified Markdown Cookbooks and local Vector RAG technologies. Furthermore, the system establishes a next-generation intelligent environment featuring a self-evolving Auto-Correction engine that tracks and learns directly from clinician adjustment patterns, all while seamlessly supporting the trusted clinical interfaces medical professionals already use.

Objective

• Implement a voice-based interface for neuronavigation registration.
• Improve the data visualization interface using the AI agent, e.g., by switching between visualization methods.
• Simplify the interface with the EEG and TMS devices AI agent.
• Minimize LLM hallucination in SlicerTMS by compiling user intent into strict template-based execution payloads.

Approach and Plan

• Zero-Hallucination RAG System: Modularize verified VTK recipes into Markdown Cookbooks, embed them into a local Vector DB, and inject real-time scene variables for deterministic execution.
• UI and UX Modernization: Integrate high-density clinical data and 2D/3D visualizations into a dynamic spatial layout to overcome Slicer’s legacy UI constraints.
• Architecture Evaluation Benchmark inference accuracy, latency, and medical data privacy between closed-network offline models and cloud-based counterparts.

Progress and Next Steps

No response

Illustrations

No response

Background and References

SlicerTMS, Slicer AI Agent, NousNav