This repository contains the code of our paper Graph Flow: Cross-layer Graph Flow Distillation for Dual Efficient Medical Image Segmentation.
The paper has been accepted in IEEE Transactions on Medical Imaging.
The paper is an extension version of our CoCo DistillNet which is published in the proceeding of the 2021 IEEE International Conference on Bioinformatics and Biomedicine (BIBM).
GastricCancer | Synapse | |
---|---|---|
Network Efficiency | Network Efficiency | Network Efficiency |
Annotation Efficiency | Annotation Efficiency | Annotation Efficiency |
|L| |
Gastric Cancer | Synapse | |||
---|---|---|---|---|---|
ACC | mIOU | average DSC | average HD | ||
|L|=1 | 0.8872 | 0.7973 | 0.7874 | 29.4551 | |
|L|=2 | 0.8874 | 0.7974 | 0.7875 | 28.7406 | |
|L|=3 | 0.8877 | 0.7980 | 0.7886 | 29.4536 |
Teacher: FANet | Hyperparameters | Gastric Cancer | ||||
---|---|---|---|---|---|---|
Student: Mobile U-Net |
λ1 | λ2 | λ3 | λ4 | ACC | mIOU |
1 | 1 | 1 | 1 | 0.7147 | 0.5560 | |
1 | 10-4 | 1 | 1 | 0.7151 | 0.5565 | |
1 | 10-9 | 1 | 1 | 0.7081 | 0.5481 | |
10-3 | 10-9 | 1 | 1 | 0.8781 | 0.7827 | |
10-5 | 10-4 | 1 | 1 | 0.7230 | 0.5661 | |
10-5 | 10-9 | 1 | 1 | 0.8800 | 0.7857 | |
10-5 | 10-9 | 0.1 | 1 | 0.8874 | 0.7974 |
test set: https://drive.google.com/drive/folders/1w2TtJBCAU0i-OQ3nb40StwzbYFBoQxQu
- Python 3.6
- Pytorch 1.7.1
- Two NVIDIA TITAN XP GPUs
The codebase of semantic segmentation is succeed from the previous work of our group.
The codebase of kd is heavily borrowed from Knowledge-Distillation-Zoo and structure_knowledge_distillation .
The pre-processed Synapse is from Transunet.
Thanks for their excellent works.