The code used for the submissions achieved an 86.85% accuracy on Private Results ranking 2nd in the in-class Kaggle competition.
The goal was to classify chest x-ray images into 3 classes:
- healthy
- bacterial pneumonia
- viral pneumonia
The training dataset consisted of 4672 images out of which:
1227 images belong to healthy subjects 2238 images belong to bacterial pneumonia subjects 1207 images belong to viral pneumonia subjects
The submissions employed:
- Data Transformation
- Data augmentation
- Transfer learning with EfficientNet and ImageNet models, trained on ImageNet
- Softmax ensemble of 16 models for final predictions