CheXNet-Pytorch

This is a binary classification(Pneumonia vs Normal) in Xray14 with Pytorch.Densenet121 is adopted directly to train a classifier,which is accessible easily in current mainstream deep learning framework,e.g. Keras,TensorFlow,PyTorch.After 160 epochs of training,I finally achieved a best accuray of 94.98%.

Dataset

The ChestX-ray14 dataset comprises 112,120 frontal-view chest X-ray images of 30,805 unique patients with 14 disease labels.I firstly extracted all normal images and imags with pneumonia,whose numbers are 1353 and 604,12 respectively.Then these images(original size is 1024*1024) are resized into 256*256 and finally I randomly split the dataset into training(80%),validataion(20%) sets.Obviously,the baises will seriously inclined to to the class with a larger number if neural network is trained roughly with raw dataset because a severely class-imbalance exists.To go further,neural network just guess all inputs as the normal and can perform a "high" accuracy(~97.81%).That's the result we don't want to see that because a terrible overfitting occures,which means such a classifier is meaningless.Data augumentation is an effective method to tackle such problems.

Data augumentation

In this project,data augumentation that is conducted for images with pneumonia is neccessary and makes great sense.The used transforms are as follows:

1. gaussain_blur:add random gaussain blur with mean=0 and variance=0.1
2. gaussain_noise:add random gaussin noise with mean=0 and variance=0.1
3. shift:randomly draft image with specified “distance”
4. rotation:randomly rotate image with specified angle
5. brightness:randomly adjust image's brightness
6. contrast:randomly adjust image's contrast

By data augumentation,the number of images increases 12 times(i.e. 13*1353=17589) compared with raw dataset.For more details,please check the script preprocessing.py.

In fact,the class-imbalance problem still exists after data augumentation.For completely eliminating such issise,I remove same normal images util the number of two classes are approximately equivalent.Hence I just randomly select 180,00 images and the left images are left out instead.Ultimately,the directory tree of processed dataset is as follows:

├── data_augu
│   ├── train
│   │   ├── LESION
│   │   └── NORMAL
│   └── val
│       ├── LESION
│       └── NORMAL

Requirement

• Python 3.6
• PyTorch 0.3

Usage

1. Clone this repository.
2. Download images of ChestX-ray14 from this released page then decompress them and finally extract all normal images and images with pneumonia into the directory NORMAL_ORIGINAL and LESION_ORIGINAL respectively.
3. Run the script preprocessing.py to accomplish data augumentation.
4. Split the entire dataset into training(80%) and validataion(20%) sets.
5. Run the scriptmain.py and train Densenet121.

Evaluation

The runtime environment is shown in the following table:

Property	Values	Note
Model	Densenet121	-
Optimizer	Adam	-
Initial learning rate	0.001	decay 0.1/40 epochs
GPUs	2*GeForce GTX 1080 Ti	-
Epochs	160	-
Mini Batch Size	50	-

Confusion matrix

Receiver Operating Characteristic(ROC) & Area Under Curve(AUC)

More

In addition,FocalLoss also with default setting is operated before data augumentation to fix class imbalance.It is a pity that there is no distinct improvement.Another training trick called Cycle Learning Rate is a kind of adjusting learning rate .Maybe it works for this project.

Before using densenet121,I try to train resnet18 but without improvement.One reason I think out may be that resnet18 has more trainable parameters than densenet121,which results in larger diffculty to train resnet18.Rather than data augumentation,fine-tune pretrained model is also a common way to do classification(Pneumonia vs Normal).If you are interesed in this idea,I recommend you to refer to this repository that obtained most stars in GitHub about this issise.I will feel great appreciated if you realize outstanding performance and make furture disscussion at your convenience because I get stuck in this aspect.

This experiment result origins from initial parameters setting without much tricks.Furture improvement are probably achieved.

Any question,please contact with me.Email(zr.estelle@gmail.com) and WeChat(zhangrong1728) are available.

---

1.For PyTorch,beside data augumentation there is another useful method called oversampling to keep class balance.However,it doesn't work in this project.You can look up the python script.