OPTIMIZING YOLOV9 AND YOLOV10 MODELS FOR BRAIN TUMOR DETECTION: A LEARNING RATE STUDY ON MRI IMAGES

Tóm tắt

This study evaluates the performance of YOLOv9 and YOLOv10 variants in detecting brain tumors in MRI images. We compared four models (YOLOv9t, YOLOv9s, YOLOv10n, YOLOv10s) while optimizing the learning rate parameter to achieve superior performance. Using the Brain_Tumor_Segmentation dataset from Roboflow containing 6,638 images divided into training (80%) and testing (20%) sets. The models were trained with hyperparameters Optimizer = SGD, lr0 = 0.00005, lr0 = 0.0001, Momentum = 0.937, Epoch = 150, Patience = 0, Batchsize = 64 and trained on Kaggle with appropriate GPU configuration. Our findings demonstrate that YOLOv10s with lr0 = 0.0001 achieves the highest overall performance with mAP(50) = 94.3%, mAP(50-95) = 72.3%, Recall = 87.3%, and Precision = 93.9%. Although the YOLOv10s model with lr0 = 0.00005 shows higher accuracy (94.2%), the increased learning rate provides a better balance between detection metrics and convergence speed.