AI Detects Cataract Severity From Pupil Images | Cheenta Research Project

Can an AI grade cataract severity from a single pupil photo? A team of Cheenta students built exactly that — an AI framework that classifies cataracts across six severity levels from real hospital images, and the hard part wasn't the model. What you'll take away from this video: ✦ How the team classified cataracts into six grades (No Cataract, NS1–NS5) from real patient pupil images → Why glare, reflections, and pupil-scale variation are the real enemies — and how partial-convolution inpainting removes them → The full pipeline: MONAI-based augmentation (rotations, blurs, zooms) feeding a ResNet-18 CNN ✦ Why combining inpainted + original images hit an ROC AUC of 0.8919 — and where the model still confuses adjacent grades Chapters: 00:00 — The problem: detecting cataracts from pupil images 01:50 — Six severity levels: No Cataract through NS5 03:45 — Core challenges: lighting, glare & pupil scale 05:53 — The ResNet-18 architecture 06:45 — Partial-convolution inpainting to remove glare 11:18 — Data augmentation with MONAI 12:35 — Results: reaching 0.8919 AUC 16:12 — Future work: interpretability & the NS2/NS3 boundary About Cheenta: Cheenta trains students not only for olympiads and entrance exams but for genuine research — the kind that turns curiosity into publishable, real-world work. Projects like this one grow out of our Research-in-School mentorship, where students tackle open problems with practising researchers. Explore research mentorship at Cheenta: → cheenta.com Related playlist: Research in School:    • Cheenta School Research   Hashtags: #CataractDetection #DeepLearning #MedicalAI #ResNet #ComputerVision #ResearchInSchool #Cheenta #MachineLearning #AIforHealthcare #StudentResearch #MONAI #ImageClassification