Doctor's Cyclop Model

Doctor's Cyclop is one project of the Phoenix team in FPT Edu Hackathon 2021. This repository contains code of the model implementation. For the Web App, view Doctors-Cyclop-Webapp

Introduction

This is an AI Model based on UNet++ helps doctors rating the placements of tubes on chest X-Rays to save lives.

This is a Model implementation integrated with Web App, the dataset is got from the RANZCR CLiP - Catheter and Line Position Challenge

After researching and referring from discussions on Kaggle, here is our model solution:

Our pipeline has 2 stages: Segmentation and Classification

Models:

Stage 1: Segmentation Model
- Using UNet++ with EfficientNet-B1 as backbone
- Train and validate on images with annotations
- For official segmentation annotation, we make 2 channel masks by drawing lines representing the tube, and drawing big dots indicating the tips of tubes.
Stage 2: Classification Model
- Using custom EfficientNet-B1 trained with noisy student
- Input has 5 channels (3 channels from original image, resized to 1024x1024 + 2 channels from predicted masks)
- Output is 12 classes: original 11 plus no-ETT, defined as whether all 3 ETT classes are 0
- Loss is weighted average of CE loss for the 4 ETT classes and BCE loss for the other 8 classes, with weight being 1:9

Training:

After 5 folds Segmentation models are all trained with original dataset, we'll generate masks for all images, including those without annotation ones.

We'll then use the predicted masks, and the images get resized to 1024x1024 as input to train stage 2 - classification models.

Then, the trained model's weights in each fold are stored in Google Drive, and will be mounted when running Online Prediction Service

How to Build and Run

1. Mount folders at Google Drive

Create shortcut (mount) to this public folder, add at your Google Drive's root (/):

After this you should now have this at your Google Drive's root:

2. SSH to server

Run this notebook with GPU enabled to create a Google Colab virtual machine
Copy file setup.sh or restapi-setup.sh to the newly created virtual machine (by scp)

3. Run

ssh to the created machine (default password is haha), then run either:
- restapi-setup.sh to start Online Prediction Server as a RestAPI service
- setup.sh for Training:
  - The default is training Classification
  - To train Segmentation instead, change is_seg to true in setup.sh

4. Note:

When training: Logs and checkpoints are automatically mounted at %Your-Google-Drive-Root%/log/
When start Online Prediction Server as a service: RestAPI is callable at /, you can ssh to this remote machine then tmux attach to get public API endpoint

thaiminhpv/Doctor-Cyclop-Hackathon-2021