MedIA-Agentic-AI

MedIA Agentic AI

This Hugging Face repository now contains both the reusable model assets and a complete source-code snapshot for the MedIA Agentic AI segmentation stack.

The GitHub repository remains the development mirror:

https://github.com/Xiang-mira/medical_agent

Contents

• teacher_models/: 23 teacher model entries and metadata.
• student_models/voxtell_style_student_round1/: reusable VoxTell-style prompt-conditioned student checkpoint.
• configs/: sanitized label, taxonomy, routing, and registry metadata.
• code/medical_agent/: complete GitHub source-code snapshot, including CLI, scheduler, wrappers, scripts, tests, docs, and lightweight third-party code.
• model_index.yaml: authoritative index of uploaded assets.
• checksums.sha256: SHA256 checksums for files in this repository.

Existing model files are preserved. The source-code snapshot was added without removing the original Hugging Face assets.

Teacher Models

The teacher assets are organized by model key:

airrc, atlasnet, atm, cads551, cads552, cads553, cads554, cads555, cads556, cads557, cads558, cads559, daps, epai_20250421, lvp, moose666, moose888, nnunet_private, saros_nnunet, totalsegmentator, unest, vista3d, vsmtrans.

For nnUNet-style teachers, each directory contains the inference metadata and selected checkpoint:

teacher_models/<model_key>/
  README.md
  inference_config.yaml
  label_mapping.json
  dataset.json
  plans.json
  fold_all/checkpoint_final.pth

Special cases:

• daps uses fold_all/checkpoint_best.pth.
• vsmtrans uses fold_0/checkpoint_final.pth.
• vista3d and unest provide MONAI-style models/model.pt assets.
• totalsegmentator is listed as an external runtime teacher with README/config only; official runtime weights are not mirrored here.
• epai_20250421 is included as an ePAI teacher checkpoint. Its public metadata intentionally omits class names; authorized lab users should use the internal label registry with the GitHub runner.

Student Model

The reusable student checkpoint is stored at:

student_models/voxtell_style_student_round1/
  README.md
  inference_config.yaml
  voxtell_finetuned_model/plans.json
  voxtell_finetuned_model/fold_0/checkpoint_final.pth
  prompt_embeddings.pt
  voxtell_prompt_train_result.sanitized.json
  voxtell_student_quality_gate.summary.json

This student is a project VoxTell-style prompt-conditioned student initialized from pretrained VoxTell assets. Do not describe it as official VoxTell fine-tuning unless the official VoxTell training entry point is explicitly used and verified.

Source Code

The executable project code is stored at:

code/medical_agent/

Important entry points:

• code/medical_agent/README.md: full project handoff and execution guide.
• code/medical_agent/run_medai_cli.py: main JSON-oriented CLI.
• code/medical_agent/agent-harness/: installable Python package and core runtime.
• code/medical_agent/scheduler/: local/Slurm resource planning and launch logic.
• code/medical_agent/scripts/: training, inference, audit, repair, and migration commands.
• code/medical_agent/tools/dataset_delivery/: 100-case and 373-target delivery tools.
• code/medical_agent/configs/: model registry, taxonomy, target space, and scheduler configs.

The source snapshot records its GitHub provenance in:

code/medical_agent/.source_commit
code/medical_agent/.source_remote

How To Use From This Repository

Clone the Hugging Face repository with LFS, then install the code from the embedded source snapshot:

git lfs install
git clone https://huggingface.co/Xiang-mira/MedIA-Agentic-AI
cd MedIA-Agentic-AI/code/medical_agent

python -m venv .venv
source .venv/bin/activate
pip install -e agent-harness

python run_medai_cli.py --json doctor
python run_medai_cli.py --json model-inventory

For real inference or training, install model-specific GPU dependencies and set the asset root to the Hugging Face repository root as needed:

export HF_ASSET_ROOT="$(cd ../.. && pwd)"

The code resolves teacher and student assets through model_index.yaml, configs/model_registry.hf.yaml, or the project configs in code/medical_agent/configs/.

Runtime Components

LabelCritic and ShapeKit are runtime/code-layer components used for auditing, comparison, and post-processing. The code snapshot includes the project-side wrappers and lightweight vendored code needed by the pipeline. Large runtime services, licensed benchmark data, private CT data, and local HPC outputs are not stored in Git history.

LabelCritic is used to compare competing candidate masks when multiple teachers predict the same structure. This is the most time- and compute-intensive part of the current workflow.

ePAI Policy

This repository includes ePAI model assets only. It does not include ePAI patch code, wrapper code, or modified export behavior. Runtime behavior belongs to the GitHub code layer.

The public ePAI metadata is redacted: class names and label semantics are distributed internally to authorized lab users, not in this public Hugging Face repository.

Safety

Research use only. These assets are not clinical diagnostic devices. Do not use them for patient care without appropriate validation and regulatory review.

Current Progress and Future Plan

Xiang is using a proof-of-concept cohort of 100 CT scans selected from about 400,000 CT scans, intended to cover nearly all 373 anatomical targets. About 20 teacher models are used; each teacher covers only part of the 373-target space, so the pipeline uses lightweight routing rather than running every model for every target.

Small-scale experiments have been completed on 20 CT scans from the 100-case cohort. Current analysis compares performance on structures that overlap with the original VoxTell label space against performance on the added non-overlapping 373-target structures. The main weak cases so far are vessel-like structures inside specific organs, which likely need post-processing.

The round-1 student is a VoxTell-style student fine-tuned on teacher-generated labels. Round 2 uses LabelCritic to choose between the round-1 teacher label and the round-1 student prediction for each structure, and round 3 repeats the same refinement setting. For round-1 to round-3 evaluation, the plan is to test pretrained VoxTell and the student models on independently annotated datasets such as TotalSegmentator and DAP Atlas.

The next step is large-scale experimentation. The main runtime cost is expected to come from teacher inference, LabelCritic comparison, and student VoxTell training/refinement across rounds.