<![CDATA[Diabetic retinopathy (DR) is a leading cause of preventable blindness, and automated detection of retinal lesions could support earlier clinical intervention if detection performance is sufficiently reliable. This paper presents a systematic, controlled comparative evaluation of five object detection models (YOLOv8n, YOLOv8s, YOLOv9c, YOLOv10n, and Faster R-CNN) on the Indian Diabetic Retinopathy Image Dataset (IDRiD) benchmark under identical preprocessing, training, and evaluation conditions. The training set (46 images, held out from the official 54 IDRiD training images before any augmentation) is expanded offline to 354 images through class-specific augmentation, including CLAHE contrast enhancement; the held-out validation images and the official 27-image test set are never augmented, preserving independence between splits. Models are evaluated on the untouched IDRiD test set using mAP50, per-class AP50, Precision, Sensitivity, F1-Score, and inference speed (FPS). Results show that YOLOv9c achieves the highest mAP50 (3.80%) with a relative strength in soft exudate (SE) detection (AP50 = 6.60%), while YOLOv8n achieves the highest sensitivity (4.22%) at real-time speed (29.5 FPS); Faster R-CNN performs poorly across all classes (mAP50 = 0.02%), indicating that two-stage detectors require domain-specific adaptation for retinal imagery. All absolute scores remain far below the level required for clinical screening, and microaneurysm (MA) detection in particular remains largely unsolved for every architecture evaluated (best AP50 = 1.27%). Rather than demonstrating clinical readiness, these findings establish a reproducible, honestly reported performance baseline and error analysis that identifies lesion scale and class imbalance, not detector choice, as the dominant bottleneck for future specialized retinal lesion detectors.]]>
Copyrights © 2026
