| Title |
|---|
| Smartphone-Based Iris Imaging and Deep Learning for Non-Invasive Diabetes Screening: A Tele-Ophthalmology Approach |
| Authors |
|---|
| Karpagam Damodaran, Bharghavy S, Gopi Krishnan, Maheswari Srinivasan, Vignesh Venkatesan, Kaja Mohideen, Abinaya B |
| Presenting |
|---|
| Karpagam Damodaran |
| PURPOSE: |
|---|
| This study aims to develop and validate a convolutional neural network–based framework for automated diabetes detection using iris image analysis focused on pancreas-related regions. |
| METHODS: |
|---|
| A prospective experimental study was conducted at Indian Diabetic Care, Chennai, involving a South Indian population aged above 20 years.High-resolution iris images were captured using a Realme Narzo 70 smartphone equipped with a SKYVIK SIGNI 20× macro mobile camera lens.Image acquisition was performed at 10 cm, with 2× digital zoom.Participants with Type I or Type II diabetes of >1 year duration were included; those with ocular surface disease, prior ocular trauma or surgery, corneal opacity, or neurological disorders affecting iris innervation were excluded. A two-stage deep learning pipeline was developed: first, a U-Net model accurately segmented the iris; then, pancreas-mapped (as per Bernard Jensen’s iridology chart) ROIs were classified using a custom CNN. |
| RESULTS: |
|---|
| A total of 395 iris image pairs from 395 patients (264 diabetic, 131 non-diabetic) were included. The dataset was split 70:30 into training (n=280; 170 diabetic, 110 non-diabetic) and test sets (n=115; 93 diabetic, 22 non-diabetic).The AI algorithm achieved a sensitivity of 93.5% (95% CI: 88.5–98.5%) and specificity of 95.5% (95% CI: 86.8–100%). The PPV and NPV values were 98.9% (95% CI: 96.7–100%)and 77.8% (95% CI: 62.1–93.5%). |
| CONCLUSIONS: |
|---|
| This study demonstrates the feasibility of AI-based iris image analysis as an effective and scalable screening approach for diabetes mellitus. |