What it does
OpticalApp detects 28 eye diseases and healthy cases using AI. It’s offline, fast, and built for low-resource settings.
Your inspiration
The project began after discovering many publicly available datasets, but noticing a lack of accessible, fine-tuned models for mobile screening. This gap sparked the idea that I could contribute something meaningful. My mother has worked in visual exams and studied visual health formally, which shaped my understanding of the problem. Initially, the app aimed to support self-examination for testicular conditions, but due to lack of datasets, I pivoted to retinal diagnostics—another area that felt personal, as I’ve experienced floaters myself. That, combined with my background and curiosity, led me to pursue AI-based eye diagnosis.
How it works
The user selects or captures a retinal image (from gallery or phone camera with an adapter). OpticalApp processes the image on-device using a lightweight EfficientNet-B0 model trained on the RFMiD dataset. The app displays the top 3 most likely outcomes (among 28 disease conditions and 1 healthy class) with their probabilities, and offers short voice descriptions in English, Spanish, and French. It requires no internet and works on basic smartphones.
Design process
After exploring other health conditions, I focused on retinal diagnostics. I trained an AI model with 29 output classes using Google Colab, which was a good middle ground for performance and cost. I used GPT as a coding assistant but designed the interface, gathered datasets, and conducted all training sessions myself. I also compared multiple models, including RETFound, but ultimately chose EfficientNet-B0 due to better real-world performance. To make diagnosis possible with a smartphone, I built a physical adapter prototype using a magnifying lens, tape, and cardboard. It cost only 10 Mexican pesos (~$0.50 USD). While I couldn’t capture clear retinal images, I was surprised by how promising and affordable the idea was. The biggest challenge was aligning the phone’s light to the pupil without blinding the subject. In future iterations, I plan to design a more stable, integrated structure with a stronger lens and better light control.
How it is different
Unlike most medical apps that require internet or expensive hardware, OpticalApp runs fully offline, supports three languages, and is designed for users with no medical background. It empowers people in underserved regions to screen their eyes without needing specialists or connectivity. The code is open source and available on GitHub so that others can try it, build upon it, or help validate it.
Future plans
- Share the APK on GitHub and enable community testing and contributions. - Continue refining the model with more diverse image data. - Improve the DIY hardware adapter into a more stable, low-cost structure. - Launch versions for iOS and web. - Add referral tools to connect users with specialists when needed. - Pursue clinical validation and certification for public healthcare use.
Share this page on