Faster MaskFormer for Coronary Artery Decease Instance Segmentation

Abstract

Coronary artery disease (CAD) remains a leading cause of global mortality, necessitating efficient diagnostic techniques. X-ray Coronary Angiography (XCA) is a commonly used imaging modality for CAD identification, but it presents challenges such as non-uniform illumination and low contrast. Medical image segmentation aids in CAD detection, where convolutional neural networks (CNNs) are commonly used. However, CNNs have limitations in modeling global relationships within data. To address this, transformer architectures, originally designed for natural language processing, have been adapted for medical imaging tasks. Inspired by recent advancements like FasterViT and MaskFormer, we introduce Faster MaskFormer, a hybrid model leveraging the strengths of both architectures. Our model exhibits superior performance in semantic and instance segmentation tasks on the ARCADE dataset compared to traditional transformer-based models like MaskFormer-Swin. Furthermore, it demonstrates reduced computational complexity and training time, making it a promising solution for CAD diagnosis.

Dataset

The ARCADE challenge dataset, namely "stenosis," which comprises 1500 annotations for stenosis detection and instance segmentation, and "syntax," which comprises another 1500 images for individual vessel classification and segmentation with 25 classes, which will also be converted to more general semantic segmentation task

/ARCADE
|
|--- train/
      |--- annotations/
      |------------train.json
      |--- images/
      |------------1.png
      |------------2.png
      ...
      
|--- validation/
|--- test/

Installation

Git clone the repository

gh repo clone VGMitkin/Faster_MaskFormer
cd Faster_MaskFormer

Create conda enviromnent

conda create -n fastermaskformer python=3.9
conda activate fastermaskformer
pip install -r requirements.txt

To run the training

• Download dataset from URL: (https://zenodo.org/records/10390295)
• Download model weights (https://drive.google.com/file/d/1UUQ0SkOnW5XC9zfg52a0l0tZYsRgEw_x/view?usp=drive_link)
• Extract files
• Config config.yaml file with your parameters. Set root to the extracted files in DATASET_ROOT argument
• Start the training by running:

python3 main.py 

• or run

sh train.sh

To run the testing

To see test results use file test_arcade_syntax.ipynb

Results Example

                                      Ground Truth                                                                       Prediction

Team

• Maxim Popov
• Vladislav Mitkin
• Arsen Abzhanov