RSNA Screening Mammography Breast Cancer Detection

6th place solution: Team Chiral Mistrals

RabotniKuma (Hiroshi Yoshihara) part

Environment

We recommend you to use Kaggle GPU docker v128.

Conda environment yaml file can be found at ./environment.yaml.

Data preparation

1. Download competition dataset and place them at ./input/rsna-breast-cancer-detection/.
2. Run image conversion script: python convert_image.py

Experiments

Expriment configs are stored in ./configs.py.

Expriment lists

Config name	Description	CV	Public LB	Private LB
Aug07lr0	Multi-view model, 1024x512	0.493	0.64	0.46
Res02lr0	Multi-view model, 1536x768	0.488	0.59	0.46
Res02mod2	Multi-view fusion model, 1536x768	0.516	-	-
Res02mod3	Multi-view fusion model, 1536x768	0.525	0.63	0.48

Run experiments

Make sure your hardware has at least a total of 48 GB of GPU RAM and run the following:

python train.py --config {config name} --num_works {number of cpu cores to be used}

Please modify batch size and learning rate in config file(./configs.py ) if your hardware has less GPU RAM.

Results (weights, predictions, training logs) will be export to ./results/{config name}/.