Pneumonia Detection from Chest X-Rays using ResNeXt50

This project applies deep learning techniques to classify chest X-ray images as Pneumonia or Normal using the ResNeXt50 architecture.

Overview

• Dataset: Publicly available chest X-ray images (Normal & Pneumonia)
• Challenge: Dataset imbalance and noisy labels
• Solution: Data resampling, augmentation, and regularization

Methodology

• Data Augmentation: Rotation, shift, brightness, zoom, etc. using ImageDataGenerator
• Model: ResNeXt50 CNN
  • Grouped convolutions with increasing filter depth
  • Regularized using L2 and Dropout
• Metrics:
  • Precision, Recall, F1-Score, Accuracy
  • Confusion matrix for detailed error analysis

Results

• Model Parameters: ~4.49M
• Performance:
  • High recall and precision for Pneumonia cases
  • Lower performance on Normal class due to class imbalance
• Training Graphs:
  • Consistent drop in training/validation loss
  • Accuracy improved over epochs

Key Takeaways

• Effective in detecting pneumonia from X-rays
• Needs improvement for detecting normal cases
• Future work: Analyze misclassifications and improve class balance