This project uses Python and TensorFlow to analyze 120,000 chest x-rays, each labeled with zero or more of 14 health issues, and identify which diagnoses apply to individual x-rays.

​

Hyperparameter Tuning - Grid search is used to find the best learning rate.

Transfer Learning - Inception Net is used as a base, and the last half of the layers are tuned on the training set. Results - Validation AUC sits at .79 right now. It's the best result on Kaggle at the moment (as of 2/22), but you still wouldn't want it to replace your doctor.

​

Click through to the GitHub repo here.

​

As a next step, I'd love to rebalance the dataset and try a mixture-of-experts model to increase precision and recall.