Inspiration
With the recent AI boom, AI model development is more prevalent than ever. However, many models being developed use large datasets and extensive computing resources. This can make it hard for people with less access to such resources to create and deploy industry standard models. In this project, we sought to investigate whether we could develop an industry standard deep learning model with resources equivalent to what is available for free online.
We wanted to do something in the medical field because it is often difficult to source large datasets, especially for independent developers. We also wanted to develop a model that could make a meaningful difference, and with Diabetic Retinopathy (DR) being one of the leading causes of preventable blindness in the world, we decided to tackle it.
What it does
Our model detects the presence and diagnoses the severity of DR in patients using fundus imaging (image of the inside of their eye).
How we built it
We fine tuned the swin-S transformer using a modest 17.5k training images for just 546 steps (2 epochs). This was done using a T4 GPU on GCP, but an equivalently free one is available on google colab for free for a limited time. Training time was under 17 minutes, enabling it to be quickly and easily trained within the free hour limit.
Log in or sign up for Devpost to join the conversation.