Pneumonia_x_ray_classification

Introduction

This project's goal is to help doctor diagnose the pathological pneumonia from chest X ray pictures using a deep learning technique.

So generally, a convolutional neural network with 4 hidden layers to classify whether an X ray image is positive of pneumonia was trained by around 5 thousands of images mixed of both pneumonia negative and positive images. The model was created with Keras Framework.

Data

Data Source:
https://www.kaggle.com/paultimothymooney/chest-xray-pneumonia
Normal example:

Pneumonia example:

Data List:

1. Training set: 5216 images
2. Dev set: 16 images
3. Testing set: 624 images

All the images are jpeg files. However, the formats of images are very varied. And some images in training set are in 'RGB' mode, while the others are in 'Greyscale' mode. Thus, all of these variations need to be consistent during the preprocessing step.

Methodology

I. Data preprocessing

1. Load images with PIL in batch by glob.
2. Convert all the images to 'RGB' mode and resize them to 128 x 128 pixels (Using larger image would cause very complex computation, and the significance change is not worth of the computation cost).
3. Standardize the pixel value and convert the image data to numpy vectors, which are the X sets for the model inputs (including training, dev and testing). Create the related Y sets (vector of 0 for normal sets, 1 for pneumonia sets) based on the X set's shape.
4. Make the complete data sets by concatenating both normal and penumonia data sets for both X and Y.
5. Shuffle the data sets in random.

II. Building the model in Keras
The general model architecture is like model VGG-16 which keeps increasing filter units in the exponential of 2 and decreasing the image size by half with max pooling during every hidden layer. After 4 hidden layers of computation, the image data is converted from (128, 128, 3) to (5, 5, 256). The flatten vector is eventually computed by a sigmoid function to get the classification results (Fig.1).

Fig.1
Hyperparameters (part of)

• optimizer: Adam, learning rate: 0.001, beta_1: 0.9, beta_2: 0.999
• epochs: 20
• mini batch size: 16

The detailed steps of data analysis can be seen in the file: X_ray_classification_v2.ipynb.

Results

By far, upon the current hyperparameters the best results are:

• Accurancy of training set: 1.0;
• Accurancy of dev set: 1.0;
• Accurancy of testing set: 0.83.

The final result can be different by more iterations of training, hyperparameters tuning or modifying model architectures.

Discussion

1. The image size can be important for the model performance. Since the difference is quite subtle for images between normal and pneumonia, larger image can be more beneficial for more details. However, the computation cost is also exponentially increased when dealing with large image. Therefore, I took the size of 128 x 128 here as the balance between performance and efficiency.
2. The sample number of dev set is quite small which is only 16. More dev samples may be good for the model performance. Because the samples were distributed by the author before shipped, no change was conducted here.
3. Other models as LeNet and AlexNet were also tried, but the performace was not significantly changed.
4. The final best accuracy of the testing set is 0.83, which is not very good and can't be reliable for actual usage. Thus, more following optimization is required definitely.

Conclusion

Using deep learning technique to assist practical clinic method is very promising and significantly useful. This model to analyze X ray images for pneumonia diagnose is a good try. However, more work needs to be done before it is really practicable.