https://github.com/hilab-git/tegda
https://github.com/hilab-git/tegda
Last synced: 7 months ago
JSON representation
- Host: GitHub
- URL: https://github.com/hilab-git/tegda
- Owner: HiLab-git
- Created: 2025-06-18T06:49:19.000Z (10 months ago)
- Default Branch: main
- Last Pushed: 2025-09-16T12:57:55.000Z (7 months ago)
- Last Synced: 2025-09-16T14:55:42.626Z (7 months ago)
- Language: Python
- Size: 12.1 MB
- Stars: 0
- Watchers: 0
- Forks: 0
- Open Issues: 0
-
Metadata Files:
- Readme: README.md
Awesome Lists containing this project
README
# TEGDA: Test-time Evaluation-Guided Dynamic Adaptation for Medical Image Segmentation
---
This is the official code for TEGDA: Test-time Evaluation-Guided Dynamic Adaptation for Medical Image Segmentation.
[2025-06] Our work have been accepted by MICCAI 2025.
## Overall Framework
Our contributions are summarized as follows:
- We present a novel prediction quality evaluation metric based on **Agreement with Dropout Inferences calibrated by Confidence (ADIC)**, where the Dice score between predictions by the model and its dropout version is leveraged to assess the robustness of the model on a testing sample, then it is further calibrated by the confidence to become highly relevant to the real Dice value between the prediction and its ground-truth
- We propose **Adaptive Feature Fusion-based Refinement (AFFR)** that adaptively fuses the feature of a sample with those with high ADIC values based on their similarity, leading to robust refined pseudo-labels.
- We introduce ADIC-guided **Self-adaptive Model Updating (SMU)** that consists of ADIC-aware pseudo-label loss weighting and ADIC-aware mean teacher to improve the stability of adaptation.
## Dataset
Download the BraTS-GLI and BraTS-PED datasets from [BraTS 2023](https://www.synapse.org/#!Synapse:syn51156910/wiki/), M&Ms datasets from [M&Ms](http://www.ub.edu/mnms).
## How to use
### Source model training
Use
```
cd code
python train_fully_supervised_2D.py # For M&Ms dataset
python train_fully_supervised_3D.py # For BraTS dataset
```
to get the source model for two datasets.
### Test-time adaptation
Use
```
./run.sh
```
to get the test-time adaptation results on two datasets.