DELAX-TB is a deep learning tool and architecture for detection of TB in x-ray scans. It's based on CheXnet and the source is based off of Bruce Chou's work at https://github.com/brucechou1983/CheXNet-Keras
Refer to the README at https://github.com/brucechou1983/CheXNet-Keras for the initial setup.
ChexNet is a deep learning algorithm that can detect and localize 14 kinds of diseases from chest X-ray images. As described in the paper, a 121-layer densely connected convolutional neural network is trained on ChestX-ray14 dataset, which contains 112,1\20 frontal view X-ray images from 30,805 unique patients. The result is so good that it surpasses the performance of practicing radiologists.
In this project:
- We took the original model from Bruce's source and loaded the saved "best" weights he shared
- Employing transfer learning, we replaced the last layer with a 3 node dense layer for tb,normal,uncertain categories of x-ray scans
- Re-trained the model with our samples
- Saved our new model
- Converted the saved model to tensorflowjs
- Used this tfjs model.json and bins for the DELAX-TB prototype
MIT