This project aims to develop an AI-powered solution for the detection of Tuberculosis (TB) using chest X-ray images. TB continues to be a major public health issue, particularly in resource-limited regions where access to radiologists is scarce. By leveraging deep learning, this tool can help automate the diagnosis process, enabling faster and more accurate detection of TB.
- Size: 5,000 chest X-ray images
- Classes: TB-positive and TB-negative
- Source: Public datasets (e.g., Kaggle Chest X-ray Dataset)
To improve generalization and robustness, several preprocessing and augmentation techniques were applied:
- Rescaling (Normalization): Scaled pixel values to [0, 1]
- Contrast Enhancement: Applied CLAHE (Contrast Limited Adaptive Histogram Equalization)
- Noise Removal: Used Gaussian Blurring to smooth images
- Edge Detection: Applied Canny edge detector to highlight features
- Image Negative: Used image inversion to analyze contrast behavior
- Homomorphic Filtering: Enhanced both contrast and illumination
- Feature Extraction: Extracted features to improve model understanding
We conducted a comparative analysis using three different CNN architectures:
- ResNet50 (Transfer Learning)
- DenseNet (Transfer Learning)
- Simple CNN (custom model)
- ResNet50 and DenseNet outperformed the Simple CNN in accuracy and AUC-ROC.
- Simple CNN had faster training times and was easier to deploy.
The repository includes:
- A Simple CNN model
- Image preprocessing pipeline
- Frontend integration to allow basic user interaction for uploading and diagnosing X-ray images
- Test Accuracy: ~99%
- Precision: 0.98
- Recall: 0.94
- F1-Score: 0.96
- AUC-ROC: 0.94
- Kaggle Chest X-ray Dataset
- ResNet & DenseNet papers
This project is a step towards using deep learning in solving real-world health issues, especially in under-resourced regions.