
Recently, I completed a project for my organization, Drishti, which performs AI-based Diabetic Retinopathy screening for rural areas in Bangladesh. To take retinal scans without full-cost retinal cameras, we built a rig that can house a smartphone to take these images. The images can then be uploaded to our research-backed AI algorithm to provide instant and accurate initial screening. This project demonstrates the harmony between software and Hardware, as we were able to effectively integrate our AI algorithm with the low-cost retinal camera rig. Hopefully you enjoy this article, and if you have any thoughts or areas of improvement, please let me know.
Software

Our algorithm is a CNN that performs Diabetic Retinopathy screening. The algorithm itself prioritizes practicality over novelty. The base network is a DenseNet-121, and the final layer is replaced with a 5-node fully-connected layer. We predict Diabetic Retinopathy classification in 5 separate stages.

For training the CNN, we used the open-source APTOS 2019 Blindness Detection fundus image dataset (APTOS). On Bangladeshi eyes, the algorithm achieves upwards of 93% accuracy, which is very impressive considering that the images it was trained on were from a separate dataset. The results confirm the potential for using the algorithm in real clinical settings in Bangladesh.

The paper with method details and results on Bangladeshi eyes is available here.
Hardware

We built our system to house a smartphone with a retinal attachment. This video demonstrates how to use it. The frame is built with PVC pipes and all the attachments are CAD designed and 3D-printable. The patient will place their chin on the chin rest and look directly into the retinal camera. A technician will then take the photo with the iPhone and upload it to Drishti’s web servers to get an instant prediction.
The system can be easily disassembled so that it can be transferred from clinic to clinic. The components are zip-tied together so that they can be taken apart if needed. Our assembly guide is here.
The cost for this build is under $20 and took about 2 months of planning, designing, and purchasing materials. The retinal attachment itself must be purchased, but the rest of the components are low-cost. The entire system works elegantly and we are currently deploying copies of the system to our two clinics that we are starting pilot runs in.
We are currently working on a v2 that adds thumb-screws to the adjustable components for easier adaptability. Our CAD is available here. If you click on the STL files, you will be able to interact with the 3D-rendered designs.
This build was done in collaboration with Sajiv Shah. He assisted in the development and design of the CAD parts and 3D-printed the materials. He also helped to assemble the first prototype and write the instructional document.
Thanks for reading this short article/blog post on the details of our project. If you have any ideas or thoughts, please let me know!