Abstract / Summary
This study aims to develop LMMR-YOLO—a lightweight deep learning-based lesion detection fusion model optimized for medical images—to efficiently and accurately identify Optical Coherence Tomography (OCT) images, measure key parameters of Macular Hole (MH), and assess MH lesion severity in patients. This retrospective study utilized 606 OCT images from 61 patients. LMMR-YOLO is built upon Mask R-CNN and YOLOv12-obb, with targeted dataset expansion via Mask R-CNN, image enhancement through the Image Enhancement Algorithm (IEA) alongside conventional augmentation ablation experiments, and YOLOv12-obb optimization including a grayscale image adaptation module integrated with Efficient Channel Attention (ECA), C2f_Ghost and A2C2f_Lite lightweight backbone modules for feature extraction, an Oriented Bounding Box (OBB) detection head, and a Fusion Detection Algorithm (FDA) that integrates Mask R-CNN and IEA for detection. Model performance was assessed using precision, recall, mAP@0.5, F1‑score and accuracy. On the internal test set, LMMR‑YOLO achieved an mAP@0.5 of 92.69%, accuracy of 94.78%, precision of 97.14%, and recall of 87.88% for key‑point detection of MH parameters, with an F1‑score of 92.28%. Normalized classification accuracy for Central Choroidal Thickness (CCT) was 0.97, mirror misidentification rate between Left Hole Height (LHH) and Right Hole Height (RHH) was 0.03, and the model featured 3.34 M trainable parameters, 11.10 GFLOPs, 9.44 FPS inference speed, and short training time, balancing lightweight performance and detection efficiency. LMMR-YOLO enables accurate MH parameter measurement from OCT images, supports ophthalmic MH diagnosis, staging, and personalized care, and offers a new paradigm for medical grayscale image analysis and model training. Not applicable.