The two-stage classifier-detector approach was followed to detect pneumonia and bounding boxes on the RSNA stage 2 dataset.
fold0-9.h5 contains pre-trained weights for detector based on RetinaNet. (10 fold CV was used to train the same) The ensemble (NMS on all predictions) of these 10 models gave AP50 of 0.518.
Remaining 4 files are pre-trained weights for the classifier. 3 neural networks and 4th one is the logistic regression model based on all the predicted probabilities. This gave a classification accuracy of about 94%