CCA-MFNet

Criss-Cross Attention based Multi-Level Fusion Network (CCA-MFNet) for Gastric Intestinal Metaplasia Segmentation

The paper is accepted by MICCAI 2022 workshop.

In this project, we demonstrate the performance of the CCA-MFNet for gastric intestinal metaplasia segmentation.

Instructions for Code:

Requirements

To install PyTorch >= 1.4.0 or 1.8.2 LTS (highly recommended), please refer to https://pytorch.org/get-started/locally/

Pytorch LTS version 1.8.2 is only supported for Python <= 3.8

It is necessary that VRAM is greater than 11GB (e.g. RTX2080Ti)

Linux Ubuntu 20.04 LTS (highly recommended)

Python 3.7

gcc (GCC) >= 4.8.5

CUDA >= 10.2

Compiling

# Install **Pytorch**
$ conda install pytorch torchvision torchaudio cudatoolkit=10.2 -c pytorch-lts

# Install **Apex**
$ git clone https://github.com/NVIDIA/apex
$ cd apex
$ pip install -v --no-cache-dir --global-option="--cpp_ext" --global-option="--cuda_ext" ./

# Install **Inplace-ABN**
$ git clone https://github.com/mapillary/inplace_abn.git
$ cd inplace_abn
$ python setup.py install

Training and Inference

Training script.

./run_train.sh

Inference script.

./run_inference.sh

Comparisons with State-of-the-art Methods (%)

Method	Backbone	mIOU	mDice	Recall	Precision	Accuracy	** Model Link**
U-Net	U-Net	57.61	65.60	61.40	77.18	94.94
U-Net++	U-Net	58.62	66.92	62.55	78.10	95.05
nnU-Net	U-Net	60.33	69.05	64.29	80.47	95.30
MedT	Transformer	53.31	59.44	56.62	71.03	94.46
W-Deeplab	Deeplab-v3	62.19	71.22	65.66	85.17	95.69
CCNet	ResNet-50	65.88	75.39	70.80	83.24	95.87
Proposed	ResNet-50	68.92	78.47	74.94	83.45	96.13	Google Drive

Reference

Please cite the following paper when you apply the code.

C.-M. Nien, E.-H. Yang, W.-L. Chang, H.-C. Cheng, and C.-R. Huang, "Criss-cross attention based multi-level fusion network for gastric intestinal metaplasia segmentation," in Proc. Int. Conf. Med. Image Comput. Comput. Assist. Intervent. Workshop Imag. Syst. GI Endoscopy, 2022, pp. 13–23.

Acknowledgments

This work was supported in part by the National Science and Technology Council, Taiwan under Grant MOST 110-2634-F-006-022, 111-2327-B-006-007, and 111-2628-E-005-007-MY3. We would like to thank National Center for High-performance Computing (NCHC) for providing computational and storage resources.

Particular Thanks

Department of Internal Medicine and Institute of Clinical Medicine, National Cheng Kung University Hospital, College of Medicine, National Cheng Kung University, Tainan, Taiwan.