Chest X-Ray Pneumonia Detector (CNN From Scratch)
Overview
A custom CNN trained entirely from scratch (random weight initialization, no pre-trained backbone) on the Kaggle Chest X-Ray Pneumonia dataset.
Architecture
- 4 Conv blocks (16→32→64→128 filters)
- BatchNorm + ReLU + MaxPool each block
- Fully connected head (256 → 2 classes)
- Input size: 128×128 RGB
- Total parameters: ~2,195,842
Dataset
- Source: Kaggle Chest X-Ray Pneumonia
- Classes: NORMAL, PNEUMONIA
- Train: 5216 images | Test: 624 images
Training
- From scratch (no transfer learning)
- Optimizer: Adam (lr=0.001)
- Loss: CrossEntropy
- Epochs: 5
- Final Test Accuracy: 69.39%
Labels
- 0 = NORMAL
- 1 = PNEUMONIA
- Downloads last month
- 17