The End
Heet D -
In this week’s post I will be going over my final part of data collection and going over my project overall.
As mentioned in the last blog post I had prepared and set up the dataset for a few-shot evaluation. Likewise I was also prepared for zero-shot evaluation based on the diseases I had picked. I began by taking the same code I had prepared for my common disease evaluation and changed the image paths and directories to match my new rare disease dataset. Then I was able to run the tests and collect the data.
Now I want to clarify my project if it has become confusing with all the different mentions of datasets and common diseases and rare diseases. Overall my main experiment is testing CLIP on the rare disease dataset (MIMIC-CXR) that I have made. But for this test to be legitimate, I need a sort of control group, or comparison baseline, which is my common disease dataset (NIH Chest X-ray). So in terms of how I actually approached this project, I first began with the common disease dataset. I began by coding a complex process to handle and manage the data for CLIP and allow it to make state-of-the-art classification on the NIH dataset. I was able to achieve this by getting a performance of 0.7506 AUC which beats the 0.742 AUC of MoCoCLIP by Bhardwaj et al. Now this indicates that my code is effective as it is on par with the code of other researchers. Once I had established this I completed the few-shot part of my common diseases and was ready to move onto the rare disease part, or my actual experiment. I hope this clears up why I took the steps that I took to obtain my data. I am now ready with my data, and this means I am essentially done with my project. Thank you!