[Multi-Modal Image+Text] Train a classifier to identify useful chest X-ray images from google images.

[sumedhasingla]

• Train a binary classifier where 1 class is NIH Chest X Ray images and 2nd class is Google Download images

• Train a binary classifier where 1 class is NIG chest X Ray images and 2nd class is natural images like from image net

• Manually assign positive and negative on google images and then train a classifier.

Test all the methods on Google downloaded images.

[sumedhasingla]

1st Try: Trained a classifer on NIG Chest XRay images vs Google downloaded images. The classifier identified the Google images with 99% accuracy. The CAM atten maps are all uniform. The expectation was, the x-ray images downloaded from Google get incorrectly classified as NIH images. But the classifier always label them as Google. So we failed.
Results: https://github.com/sumedhasingla/MultiModalImageText/blob/master/Explore_Predictions_Binary_Classification_NIH_vs_Google.ipynb