부산대학교 도서관

Automatic landmark detection in knee x-rays requires an optimizer to improve its performance while shortening training time. The aim of this research is to find a suitable optimizer for the YOLOv8m model for the task. We focus the comparison on six optimizers, including SGD, Adam, AdaMax, AdamW, NAdam, and RAdam. This study used 400 normal Kellgren-Lawrence (KL-0) right knee x-ray images from the Public Osteoarthritis Initiative (OAI) dataset with 5.600 landmark spatial features. Each knee x-ray image in the dataset is labeled with 14 landmarks on the femur and tibia bones. We resized the knee x-ray image as input to a size of 640×640 pixels, then using the YOLOv8m model with batch size 16 and epoch 150, we carried out training and testing with the SGD, Adam, AdaMax, AdamW, NAdam, and RAdam optimization algorithms one by one. We found that the optimizer average precision (mAP) was 0.696 with the Adam optimizer outperforming the others, while the precision of 0.716 and recall of 0.685 with the AdaMax optimizer outperformed the others, and the fastest time was 0.385 hours with the AdamW optimizer.