A grade is a number.
Understand what it means.
Upload a retinal photograph to run the trained grader, then ask the live assistant about your result. You can also explore illustrative examples.
Talk through the result
A research demonstration. Responses combine a live Qwen language model with explicit scope guards, supplied grading facts, and cited category definitions. They do not confirm image findings or determine medication.
What is behind this demo?
The grading study
10,000 EyePACS training images, 2,910 APTOS validation images and 1,636 Messidor-2 test images, matched across retinal evidence streams.
The frozen ResNet-152 checkpoint is used for uploads. An audit found different X3 preprocessing in the previous training and external-evaluation scripts. Historical scores do not establish performance for this upload pipeline; consistent reevaluation is required.
The language assistant
Qwen2.5-1.5B-Instruct plans optional explanation details. Supplied facts and scope guards control the response. The standalone LLM routed 24/36 development examples correctly; the guarded flow passed 36/36. This is not an independent clinical benchmark.
Uploaded photographs run through the original X1–X4 feature extractors and frozen ResNet-152 grader. The assistant explains the server-held prediction. Example contexts remain clearly illustrative. No clinical validation or treatment decisions are provided.