Mirrored X-Net

This repository contains pytorch supported code and configuration of the proposed weakly supervised model for geographic atrophy (GA) lesions segmentation:

Contents

• Introduction
• Dataset
• Running
• Requirements

Introduction

• TO DO: Introduce our work.

Dataset

We utlize two different GA datasets (Dataset 1 and Dataset 2) and one normal (Dataset 3) dataset to evaluate the model. Each dataset contains SD-OCT cubes with three dimensions. All the cubes contain the advanced non-exudative AMD with GA.

Our datasets were acquired with a Cirrus OCT device from Stanford University.

The summarizations of data information on three datasets

	Dataset 1	Dataset 2	Dataset 3
Size	512×128×1024	200×200×1024	512×128×1024
Cubes	51	54	63
Individuals	8	54	38
Eyes	12	54	62
Category	GA	GA	Normal

Running

Data preparation

• Our file structure looks like:

$ tree data
├── dataset
│   ├── CubeScan
│   │   ├── eye1_cube_z
│   │   │   ├── 1.bmp
│   │   │   ├── 2.bmp
│   │   │   ├── 3.bmp
│   │   ├── eye2_cube_z
│   │   ├── eye3_cube_z
│   ├── SegmentationBScan
│   │   ├── eye1_cube_z
│   │   │   ├── 1.bmp
│   │   │   ├── 2.bmp
│   │   │   ├── 3.bmp
│   │   ├── eye2_cube_z
│   │   ├── eye3_cube_z
│   ├── Segmentation2D
│   │   ├── eye1_cube_z.bmp
│   │   ├── eye2_cube_z.bmp
│   │   ├── eye3_cube_z.bmp
│   ├── statistics.xls

Notes:

• Especially, GA is defined as a 2D lesion manifestation. The training datas are 3D cubes, the ground truths are 2D projections. We use B-scans in random order as input for training phase, all B-scans of a cube as input for test phase to produce a projection attention map.
• Normal datasets, e.g. Dataset 3, can be introduced as additional normal images.

Configurations

• The codes are follow pythorch

  • The path of model building is models\
  • The path of dataset building is datasets\
  • The running file is main.py

• The hyper-parameter formats follow mmcv.

  • The path of model configuration is configs\models\mxnet.py.
  • The paths of dataset configuration are configs\datasets\GA_128_bscan.py and configs\datasets\GA_200_bscan.py.
  • The paths of cross validation configuration are configs\cross_validation_128\ and configs\cross_validation_200\.

Training and Testing

# single-gpu running
CUDA_VISIBLE_DEVICE=0 python main.py <CONFIG_FILE> --work-dir <WORK_DIR> --show-dir <SHOW_DIR>

# example for one fold of cross-validation
CUDA_VISIBLE_DEVICE=0 python main.py conifgs/cross_validation_128/1.py --work-dir output/128/1 --show-dir ./show/128/1

Notes:

• This version of code does not support distributed training and testing for now.
• This version of code does not support loading preatrained parameters for now.

Requirements

• My running environments, based on Ubuntu 18.04:
  • torch
  • mmcv-full
  • timm
  • loguru
  • skimage
  • pydensecrf

Citing

If this project is help for you, please cite it.

@article{ji2024mirrored,
  title={Mirrored X-Net: Joint classification and contrastive learning for weakly supervised GA segmentation in SD-OCT},
  author={Ji, Zexuan and Ma, Xiao and Leng, Theodore and Rubin, Daniel L and Chen, Qiang},
  journal={Pattern Recognition},
  volume={153},
  pages={110507},
  year={2024},
  publisher={Elsevier}
}