Korea Health Datathon 2020 Sinusitis Classification Solution

This repository is the 3rd place solution the sinusitis classification of KHD2020 based on sample code.

NOTE : It is unable to run this code on your local machine because the challenge was based on NSML infrastructure and the dataset was private. You have to customize the data_loader.py to run training code with your own dataset and specify the directory on DATASET_PATH argument.

Getting Started

Requirements : run pip install -r requirements.txt

Pre-processing : Check notebook to see specific pre-processing process.

1. Zero-padding to 300x600
2. Windowing
3. Background reduction
4. RoI crop to 224x224
5. Min-Max scaling

Training Details :

We trained network using SGD optimizer with a momentum of 0.9 and decay of 0.3. However, we didn't apply weight decay on bias term. We used a decaying learning rate with a cosine annealing warm-up start method setting an initial learning rate to 0.0005 and a minimum rate as 5e-6. We used a small batch size of 8 to increase training stability and trained the network for 60 epochs. The sinusitis dataset has a severe class imbalance, therefore, we adopted class weights for loss function with the ratio 1:4:6:9. To avoid overfitting, we added dropout on the fully connected layer with 0.5 probability. We applied simple data augmentation techniques such as random rotation (-15\~15 degress) and scaling (x0.85\~1.15).

Structure :

• model/: network architecture definitions and training/validation core functions
• utils/: image processing code, config argument parser, and data loading code
• main_pytorch.py: code for model training and validation

NSML Environment

• Main command : nsml run -d 2020KHD_PNS -e main_pytorch.py -c 8 -a '--batch_size 64'
• Submit command : nsml submit NSML_ID/2020KHD_PNS/SESSION_NUM CHECKPOINT_NUM

Local Environment

• Main command : python main_pytorch.py --DATASET_PATH YOUR_DATASET_DIR --batch_size 64