Convolutional NN: Pneumonia Detection

Visit my blog for a more in depth tutorial on how to create a Machine Learning/AI that predicts pneumonia given a lung x-ray image.

Data

1. Kaggle CoronaHack
2. GitHub COVID Dataset
3. GoogleDrive(contains combined datasets, pickles, and csv with image paths/labels)

CNN Layers/Architecture

Code

Train Test Split

All of the train, test, and pickle files can be downloaded from my GoogleDrive

After splitting the train and test set, the class frequencies were as follows:

Code

Loss and Validation

The model began overfitting at about 30 epochs and had F1, AUC, and ROC scores of .935, .972, and .994 respectively. Code

Threshold

Code

Confusion Matrix for three thresholds

Code

Test on Google Images

To test the model on outside data, I randomly gathered 17 images from Google and used a probability threshold of .65. As you can see below, the model had a 100% True Positive rate and 56% True Negative Rate. If you download the code in my repo, there is a folder where you can try a prediction yourself. All you have to do is download the image, name the file normal or non-normal to remember its label, then run the code

Future Directions

• Classify the type of Pneumonia (Viral, Fungal, etc.) 
• Use grey-scale (1D array)
• Apply SMOTE to alter class imbalances 
• Use larger images with dimensions of (96,96,3) or (204, 204,3). I initially tried using larger dimensions but my computer was too low. Having larger images can make detection more accurate as there is more detail than the latter.