Digital Medicine 2020 Fall: RSNA ICH Detection

[slide]

Task

Design and present an analysis flow for RSNA ICH classifiers.

Environment

• Ubuntu 20.04.1 LTS
• Python 3.7.5+
• GeForce GTX 1080 Ti
• CUDA version 11.0+

Packages

• tqdm
• numpy
• pandas
• pydicom
• seaborn
• openpyxl
• matplotlib
• scikit-learn 0.23.2+
• pytorch 1.7.0+
• torchvision 0.8.0+

Prerequsites

pip3 install -r requirements.txt

Usage

Training example

python3 train.py \
--exp_name='' \
--model_name googlenet \
--generate_exp_name --name_split_file --pretrained --radam \
--random_apply_aug --random_horizontal_flip --random_rotation --random_erasing --random_order \
--cutmix

Testing example

python3 test.py \
--model_name googlenet \
--exp_name='_Pretrain_googlenet_E100_lr0.0005_b1_0.9_b2_0.999_bs_16_splt_0.7_prekeral_bsb_radam_cutmixbeta_1_cutmixprob_0.5_rand_app_rand_flip_rand_rota_rand_ord_rand_eras' \
--cpt_name='googlenet_E_79_iter_21040.cpt'

Supported models

• resnet18, resnet34, resnet50, resnet101, resnet152
• resnext50_32x4d, resnext101_32x8d
• wide_resnet50_2, wide_resnet101_2
• densenet121, densenet169, densenet161, densenet201
• googlenet

Our method

Data preprocessing

• Brain/Subdural/Bone Window (https://www.kaggle.com/reppic/gradient-sigmoid-windowing)

Data Augmentation

• Random Apply
• Random Order
• Random Horizontal Flip
• Random Rotation
• Random Erasing / Cutout
  • Zhong, Zhun, et al. "Random Erasing Data Augmentation." AAAI. 2020.
  • DeVries, Terrance et al. "Improved regularization of convolutional neural networks with cutout." arXiv preprint 2017.
• Mixup
  • Zhang, Hongyi, et al. "mixup: Beyond Empirical Risk Minimization." ICLR 2018
• Cutmix
  • Yun, Sangdoo, et al. "Cutmix: Regularization strategy to train strong classifiers with localizable features." ICCV. 2019.

Model

Fine-tune models in torchvision.models with ImageNet pretrained weight

• Simply replace last fully-connected layer to fit our data
• RAdam optimizer (https://github.com/LiyuanLucasLiu/RAdam)

Experiment results

https://docs.google.com/spreadsheets/d/1HslC0VL4noqPN9LpZZlKMbwCwXAd38pN6IiANSGjKDk/edit?usp=sharing