Pneumonia-Classification
Pneumonia Detection using CNN based classification
Pneumonia Detection using CNN based classification model trained on chest x-ray scans of normal and people diagnosed with pneumonia. We trained a Convolutional Neural Network designed to process these x-ray images of the chest with pre-classified labels from the training set to replicate those results on the test set.
Data Visualization & Preprocessing
We trained on a dataset of:
- over 5216 total x-ray images
- 1341 x-ray images of normal class
- 3875 x-ray images of pneumonia class
- highly imbalanced
We tested on a dataset of:
- over 624 total x-ray images
- 234 x-ray images of normal class
- 390 x-ray images of pneumonia class
For more info on the data: https://data.mendeley.com/datasets/rscbjbr9sj/2
Images from both the classes
Visualizing the data
As you can clearly see the data is imbalanced. Training a model on this imbalanced data would result in naive behaviour where the model would be always favoring the pneumonia class and still produce a decent accuracy but such results would be useless. To avoid this overfitting, we will increase the number of training examples using data augmentation.
Model details
| Layer (type) | Output Shape | Param |
|---|---|---|
| conv2d-5 (Conv2D) | (150, 150, 32) | 320 |
| max-pooling2d-5 (MaxPooling2 | (75, 75, 32) | 0 |
| conv2d-6 (Conv2D) | (75, 75, 64) | 18496 |
| max-pooling2d-6 (MaxPooling2 | (38, 38, 64) | 0 |
| conv2d-7 (Conv2D) | (38, 38, 128) | 73856 |
| max-pooling2d-7 (MaxPooling2 | (19, 19, 128) | 0 |
| conv2d-8 (Conv2D) | (19, 19, 128) | 147584 |
| max-pooling2d-8 (MaxPooling2 | (10, 10, 128) | 0 |
| flatten-1 (Flatten) | (12800) | 0 |
| dropout-4 (Dropout) | (12800) | 0 |
| dense-2 (Dense) | (512) | 6554112 |
| dense-3 (Dense) | (1) | 513 |
Total params: 6,794,881
Trainable params: 6,794,881
Non-trainable params: 0
Inputs and output
- The input is a batch of images and each image has a shape (150,150,1).
- The output represents a binary classification of the input images as 1 (Pneumonia) or 0 (Normal).
Results
On the test set, we achieved:
- Accuracy = 0.92
- F1-score = 0.92
- Recall = 0.93
- Precision = 0.91
Model type conversion
The model was converted into a TFLite model using the TFLiteConverter for deploying it to android Apps.
Stay Healthy App
- The app uses the generated tflite model for image classification.
- User Uploads the chest X-ray of the person.
- The image is first converted into a bitmap and then each pixel is extracted and stored into a ByteBuffer.
- An Interpreter object is instantiated which holds the tflite model.
- This interpreter object is then invoked by passing 2 ByteBuffer objects (the input and the output), the output is modified during this call.
App Permissions
- READ EXTERNAL STORAGE
Screenshots
