|
|
--- |
|
|
library_name: pytorch |
|
|
license: apache-2.0 |
|
|
tags: |
|
|
- medical-image-segmentation |
|
|
- interactive-segmentation |
|
|
- nnunet |
|
|
- meningioma |
|
|
- radiotherapy |
|
|
pipeline_tag: image-segmentation |
|
|
--- |
|
|
|
|
|
# Interactive-MEN-RT |
|
|
|
|
|
**Domain-Specialized Interactive Segmentation for Meningioma Radiotherapy Planning** |
|
|
|
|
|
**Status**: Research prototype — Not for clinical use. |
|
|
|
|
|
## Overview |
|
|
|
|
|
Interactive 3D meningioma segmentation framework built on nnU-Net v2 and nnInteractive. |
|
|
|
|
|
- **Performance**: 77.6% Dice, 64.8% IoU (BraTS 2025 Meningioma RT) |
|
|
- **Interaction Modes**: Point, scribble, box, lasso |
|
|
- **Input**: T1c MRI (contrast-enhanced) |
|
|
|
|
|
## Files |
|
|
``` |
|
|
nnUNetInteractionTrainer__nnUNetPlans__3d_fullres_scratch/ |
|
|
├── plans.json |
|
|
├── dataset.json |
|
|
└── fold_0/ |
|
|
├── checkpoint_best.pth (820 MB) |
|
|
└── checkpoint_final.pth (820 MB) |
|
|
``` |
|
|
|
|
|
## Quick Start |
|
|
```python |
|
|
from huggingface_hub import snapshot_download |
|
|
import os |
|
|
|
|
|
# Download checkpoint |
|
|
root = snapshot_download( |
|
|
"hanjang/Interactive-MEN-RT", |
|
|
allow_patterns=["nnUNetInteractionTrainer__nnUNetPlans__3d_fullres_scratch/**"] |
|
|
) |
|
|
CKPT = os.path.join(root, "nnUNetInteractionTrainer__nnUNetPlans__3d_fullres_scratch") |
|
|
|
|
|
# Load and run inference |
|
|
from Interactive_MEN_RT_predictor import InteractiveMENRTPredictor |
|
|
import torch, numpy as np |
|
|
|
|
|
predictor = InteractiveMENRTPredictor( |
|
|
device=torch.device("cuda" if torch.cuda.is_available() else "cpu") |
|
|
) |
|
|
predictor.initialize_from_trained_model_folder( |
|
|
model_training_output_dir=CKPT, |
|
|
use_fold=0, |
|
|
checkpoint_name="checkpoint_best.pth" |
|
|
) |
|
|
|
|
|
# Run on your volume (shape: 1, H, W, D) |
|
|
predictor.reset_interactions() |
|
|
predictor.set_image(volume) |
|
|
predictor.set_target_buffer(np.zeros_like(volume[0], np.float32)) |
|
|
predictor._finish_preprocessing_and_initialize_interactions() |
|
|
predictor._predict_without_interaction() |
|
|
prediction = (predictor.target_buffer > 0.5).astype(np.uint8) |
|
|
``` |
|
|
|
|
|
# Citation |
|
|
|
|
|
```bibtex |
|
|
@inproceedings{interactive-men-rt-2025, |
|
|
title={Domain-Specialized Interactive Segmentation Framework for Meningioma Radiotherapy Planning}, |
|
|
author={Junhyeok Lee, Han Jang and Kyu Sung Choi}, |
|
|
booktitle={MICCAI CLIP Workshop}, |
|
|
year={2025} |
|
|
} |
|
|
``` |
|
|
|
|
|
# Links |
|
|
|
|
|
GitHub: [snuh-rad-aicon/Interactive-MEN-RT](https://github.com/snuh-rad-aicon/Interactive-MEN-RT) |
|
|
|
|
|
Contact: janghan001112@gmail.com |
|
|
|
|
|
Developed at Seoul National University AICON Lab |
|
|
|
|
|
Research only. Not for clinical use. |
|
|
|
|
|
--- |
|
|
|
