Diabetic retinopathy (DR) is a major contributor to blindness worldwide, particularly in areas with limited access to ophthalmic services. Timely identification is critical to preventing permanent vision impairment and improving clinical outcomes. This research introduces a lightweight, real-time Assisted Mobile Diagnostic (AMD) platform for DR screening that embeds a refined deep learning model directly onto a mobile device. Unlike traditional AI solutions that rely on powerful servers and fixed fundus imaging equipment, the proposed setup pairs a non-mydriatic retinal camera with a smartphone, enabling on-site evaluations. Retinal photographs undergo initial processing steps, including smoothing and contrast adjustment, before classification by a fine-tuned DenseNet-121 network. Model training leveraged a combination of an internal hospital dataset and two prominent public repositories: APTOS (Asia Pacific Tele-Ophthalmology Society) and EyePACS (Eye Picture Archive Communication System). The method performed strongly, achieving 97.38% accuracy on APTOS, 90.90% on EyePACS, and 98.61% on the private dataset. On mobile hardware, the system processes images in an average of 162.5 ms, enabling efficient point-of-care screening. This deep learning-powered mobile tool forms a core component of a broader tele-ophthalmology ecosystem encompassing tablet and desktop versions, promoting widespread and remote access to DR diagnosis.
