https://github.com/werserk/aiijc-med-1st-place

Solution of the first place in the AIIJC competition for detecting signs of coronavirus pneumonia on CT scan of the lungs
https://github.com/werserk/aiijc-med-1st-place

competition computer-vision covid19 deep-learning hackathon opencv segmentation torch

Last synced: 3 days ago
JSON representation

Solution of the first place in the AIIJC competition for detecting signs of coronavirus pneumonia on CT scan of the lungs

• Host: GitHub
• URL: https://github.com/werserk/aiijc-med-1st-place
• Owner: werserk
• License: mit
• Created: 2022-01-13T17:23:27.000Z (almost 3 years ago)
• Default Branch: main
• Last Pushed: 2022-12-25T10:17:03.000Z (almost 2 years ago)
• Last Synced: 2023-03-05T09:11:47.304Z (over 1 year ago)
• Topics: competition, computer-vision, covid19, deep-learning, hackathon, opencv, segmentation, torch
• Language: Python
• Homepage: https://aiijc.com/en/
• Size: 2.1 MB
• Stars: 0
• Watchers: 1
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md

Awesome Lists containing this project

README

        
# AIIJC. The "AI in Medicine" competition track

## Competiton overview

Artifical Intelligence International Junior Contest (AIIJC) has 10 tracks on different themes. \

This repository contains the code for the "AI in Medicine" track. \

Specific of competition is that all participants must be under 18 years old.

The main organizer and sponsor of the competition is Sberbank. \

Organizer and task developer is Alliance for Artificial Intelligence.

## Data

Competition don't provide any data. \

So we find 3 datasets for training and testing our models:

* [MosMedData](https://mosmed.ai/datasets/covid19_1110/)

* [MedSeg](http://medicalsegmentation.com/covid19/)

* [Zenodo](https://zenodo.org/record/3757476#.YRqU0IgzbP_)

## Methods

### Pathology segmentation

For this task we used:

* 2 pretrained Unet++ with backbone EfficientNet-B7 for 2 classes: \

  COVID-19 ground glass and COVID-19 consolidation.

* Loss function is IoU

* Optimizer is Adam with learning rate 1e-4

### Lungs segmentation

Our task need percentage of pathologies per lung. \

So we need to segment lungs. \

For this task we prepared specific computer vision pipeline.

1. The mask of pathology was subtracted from the image

   We did it because pathology prevents

   us from seeing the lungs (they are totally black, pathology - gray/white).

2. Find 2 largest contours in the image

   The contours are the lungs.

3. Delete image by half and find max overlaps for each contour with half of the image.

   Here we find the lungs that are on the left and right. (in dicom we have info about the side of the image)

4. Cut mask of pathology by contours of lungs

   Our model isn't perfect, so we need to cut the mask of pathology by contours of lungs.

   This was we reached better accuracy.

5. Calculate percentage of pathology per lung

Let's see how it works in images:

...

## Results

![](.images/website.png)

|            image             |      base opencv adaptive thresholding      |         ours adaptive thresholding          |

|:----------------------------:|:-------------------------------------------:|:-------------------------------------------:|

| ![](.images/slice_image.png) | ![](.images/base_adaptive_thresholding.png) | ![](.images/ours_adaptive_thresholding.png) |

## Run code

Clone this repository

```bash

git clone https://github.com/werserk/AIIJC-med-1st-place.git

```

Install folder with model checkpoints: [link](https://drive.google.com/file/d/19svztOBB4RhnW7cwuZTDPZb0EiWKdydN/view?usp=sharing)(could be unavailable, contact me)

```bash

pip install -r requirements.txt

```

To activate web interface run

```bash

streamlit run streamlit_app.py

```

To activate app in bash run

```bash

python3 bash_app.py --data  --save_folder  --multi --show_legend

```