doublechenching/xray14-keras

training a classification model with xray14 dataset

xray14 thoracic disease classification

Implement and evaluation of xception keras model with xray14 dataset. After 23 epochs training with vanilla xception model ,I get a roc score close to current second score.

example of xray14

vanilla xception model roc score

1. roc score comparing with known publish result
2. roc curve

problems I found

1. offical train_val and test dataset splitting have a heavy bias.when i plot 14 diseases distribution,

there is a heavy unbias.There are total 86523 images in train and val dataset and 25595 images in test dataset, dataset spliting ratio close to 3:1, but disease splited ratio close to 2:1.Thus, if you train with this splitting without any weighted loss function, you will find generalization performance drop sharply.
In experiment, I used two blocks of attention model integrated in xception model. I got lower score, but still close to naive xception model, I get average roc score is 0.87 in validation daset.I think my mode is sota. But thing is more than that.When I evaluate test dataset, there is 0.7 roc score decrease. Maybe my model is not rubust enough, but dataset spliting is really problem. I think i will get better result if I split dataset in patient level making distribution more blanced.

14 disease distribution in full dataset

There are total 112120 images. Infiltration ---19894
Effusion ---13317
Atelectasis ---11559
Nodule ---6331
Mass ---5782
Pneumothorax ---5302
Consolidation ---4667
Pleural_Thickening ---3385
Cardiomegaly ---2776
Emphysema ---2516
Edema ---2303
Fibrosis ---1686
Pneumonia ---1431
Hernia ---227

14 disease distribution in train val dataset

There are total 86523 images.
Infiltration ---13782
Effusion ---8659
Atelectasis ---8280
Nodule ---4708
Mass ---4033
Consolidation ---2852
Pneumothorax ---2637
Pleural_Thickening ---2242
Cardiomegaly ---1707
Emphysema ---1423
Edema ---1378
Fibrosis ---1251
Pneumonia ---875
Hernia ---141

14 disease distribution in test dataset

There are total 25595 images.
Infiltration ---6112
Effusion ---4658
Atelectasis ---3279
Pneumothorax ---2665
Consolidation ---1815
Mass ---1748
Nodule ---1623
Pleural_Thickening ---1143
Emphysema ---1093
Cardiomegaly ---1069
Edema ---925
Pneumonia ---555
Fibrosis ---435
Hernia ---86

2. attention model focus wired feature

In experiment, i use two blocks of attention model, but found saliency feature is so wired and i still got decent roc sorce

Saliency map