Implementation of a Deep Learning Model for Real-time Detection of Diabetic Retinopathy in Primary Care Clinics

The worldwide prevalence of diabetic retinopathy (DR) causes blindness among preventable conditions even though screening guidelines exist. This research project developed and tested a deep learning detection system for DR in real-time which operated across five different types of primary care medical facilities to resolve the current discrepancies between AI design and clinical application. The system utilized an EfficientNet-B4 architecture with attention mechanisms for training 100.000 fundus images which were then integrated with standard retinal cameras on edge computing devices. The prospective 2,500 patient diabetic study achieved diagnostic results that matched expert assessment through sensitivity 94.7 % and specificity 93.2 % and Area under the Curve value of 0.967. When deployed in practice the system achieved a 67 % rise in screening adherence as well as an 78 % reduction in patient wait times and a 43 % decrease in per-person screening expenses. Qualitative findings revealed two essential implementation factors which consisted of tailored workflow processing alongside extensive provider technical training sessions. Research shows how deep learning-based DR screening can operate successfully in primary care facilities at a reasonable cost while supporting effective clinical outcomes to shift diabetic eye care from specialist to primary care provider management.

Перевод пока недоступен