AI / ML · Research
Eye Disease Detection
A peer-reviewed study, published in IJERT (Vol. 14, Issue 05, 2026), on classifying eye diseases from retinal imagery using transfer learning. We compared three convolutional architectures, VGG19, EfficientNetB0, and MobileNetV2, with accessible, low-cost medical screening as the goal. EfficientNetB0 reached 94.2% validation accuracy across four classes.
- Role
- ML research & engineering
- Client
- IJERT (peer-reviewed journal)
- Timeline
- Published May 2026
- Year
- 2026
EfficientNetB0: 94.2% validation accuracy across 4 classes
Cataract, diabetic retinopathy, glaucoma, and normal fundus images
Published in IJERT, DOI 10.17577/IJERTCONV14IS050007
Overview
Published as a peer-reviewed paper with co-authors, this project investigates how well modern convolutional networks can classify four eye conditions (cataract, diabetic retinopathy, glaucoma, and normal) directly from retinal fundus images. The motivation is practical: where a specialist is not nearby, a reliable screening model can help triage who needs urgent care.
The challenge
Medical imaging datasets are small, imbalanced, and noisy compared with the datasets these networks were originally trained on. Training from scratch would overfit, so the real question was how far transfer learning could carry compact, deployable models.
Approach
I built a pipeline around transfer learning, fine-tuning VGG19, EfficientNetB0, and MobileNetV2 on the retinal dataset and comparing them across accuracy and generalisation. The work covered the full loop: preprocessing and augmentation, training, and evaluation, with care taken not to over-read a single metric.
Outcome
The comparison gives a clear, honest picture of the trade-offs between a heavy classic architecture and leaner, efficiency-first ones for this task, the grounding needed before anything approaches real screening use. EfficientNetB0 came out ahead, with a peak training accuracy of 98.5% and 94.2% on validation, while VGG19 and MobileNetV2 traded accuracy for size and speed. The paper was published open access in the International Journal of Engineering Research & Technology (IJERT), Vol. 14, Issue 05 (IIRA 5.0, 2026), DOI 10.17577/IJERTCONV14IS050007.
Key features
Transfer learning across VGG19, EfficientNetB0, and MobileNetV2
Preprocessing and augmentation tuned for small, imbalanced medical datasets
Head-to-head evaluation on accuracy and generalisation, not a single metric
A leaner, deployable model path aimed at real screening use
Built with
Have a project like this?
I take products from first sketch to shipped, design, build, and the AI in between. Tell me what you’re working on.
