liyu10000/pneumoconiosis

Deep Learning Exploration of ChestXRay Dataset

CheXNet Exploration

Source of the ChestX-ray14 paper and data

• original paper: https://arxiv.org/pdf/1705.02315.pdf
• paper by NG: https://arxiv.org/abs/1711.05225
• dataset_nih: https://nihcc.app.box.com/v/ChestXray-NIHCC
• dataset_kaggle: https://www.kaggle.com/nih-chest-xrays/data

Analysis of dataset

• concerns from a radiologist: Exploring the ChestXray14 dataset: problems

Model implementation

Since the models described in source paper and paper by NG's group are not available, I did some search and located some useful dataset analyze and model exploration efforts made by people on Kaggle and Github.

• Kevin Mader on Kaggle. He has presented very good dataset analysis, data preprocessings and model training.
• Caleb P on Kaggle. Tried to MobileNet and InceptionResNetV2 as base model and showed sample data training and result presenting.
• arnoweng on GitHub. It is a pytorch reimplementation of CheXNet, that presented by NG paper. It only has prediction code, not training code.
• brucechou1983 on GitHub. It is a keras reimplementation of CheXNet. Contains full codes.

My exploration

• cnn1, cnn2: Used the data preprocessing and model building method described by Kevin Mader.
• cnn3: Used data preprocessing method by Kevin Mader and model building method by Caleb P.
• cnn4: Used data preprocessing method by Kevin Mader and model building method by brucechou1983. Updated image proprocessing method: center cropping, 0-1 normalization, mean/std normalization. Train on 2 classes.
• cnn5: Used data preprocessing method by Kevin Mader and model building method by brucechou1983. Updated image proprocessing method: center cropping, 0-1 normalization, mean/std normalization. Train on 14 classes.
• cnn6: Used data preprocessing method and model building method by brucechou1983. Followed the styles of optimizer, learningratescheduler, class_weights. Train on 14 classes.
• cnn7: Used data preprocessing method and model building method by brucechou1983. Followed the styles of optimizer, learningratescheduler, class_weights. Train on 2 classes.
• cnn8: Used InceptionV3 as base model, train on 2 classes.

Results

• per-class AUROC value of model trained with bruce's code, in CheXNet-Keras/experiments/1.

Pathology	Wang et al.	Yao et al.	CheXNet	Our Model
Atelectasis	0.716	0.772	0.8094	0.7810
Cardiomegaly	0.807	0.904	0.9248	0.8694
Effusion	0.784	0.859	0.8638	0.8690
Infiltration	0.609	0.695	0.7345	0.7066
Mass	0.706	0.792	0.8676	0.7322
Nodule	0.671	0.717	0.7802	0.6123
Pneumonia	0.633	0.713	0.7680	0.6867
Pneumothorax	0.806	0.841	0.8887	0.8224
Consolidation	0.708	0.788	0.7901	0.7242
Edema	0.835	0.882	0.8878	0.8918
Emphysema	0.815	0.829	0.9371	0.8743
Fibrosis	0.769	0.767	0.8047	0.7158
Pleural Thickening	0.708	0.765	0.8062	0.7725
Hernia	0.767	0.914	0.9164	0.8020

Useful links

• keras checkpoint saving: https://machinelearningmastery.com/check-point-deep-learning-models-keras/