This project focuses on Hierarchical thoracic disease detection. It works in three layers each specialised in specific classification to narrow down a specific disease.The first layer trained on RAN (Residaul attention Network) to first identfy if the image provided has an abnoramlity or not. If yes, the information about this is passed on to the second layer based on (CheXnet Localisation), which identifies, in which area of the image there is an abnormality. Finally, once the location of this abnormality is pinpointed, layer three the local and global features, to present a final verdict of the disease Not only is this method significantly more accurate, due to its segregation and specialization approach, it is also considerably cost efficient to train and use.

The main file to run the project in the Project Main/Entire pipeline.ipnyb file. It references pretrained models, so make sure you have them in the right locations.

Classification Hierarchy:

Images of the result:

A PDF about the final presenation along with data insights, proposed model architecture, experiment results, and inferences can also be found attached.

Feel free to get in touch about the project with us at : [email protected] [email protected] [email protected] [email protected]