Comparative Analysis of Optimizers in Lumbar Spine MRI Segmentation Using ResNet50-based U-Net

In this study, a ResNet50-based U-Net model was utilized for lumbar spine segmentation, and the effects of different optimization algorithms were thoroughly analyzed. Various optimization techniques, including Adam, AdamW, Adamax, RMSprop, Adadelta, Adagrad, and Adafactor, were tested to enhance the learning process and overall segmentation performance on MRI images. The model’s performance was evaluated using metrics such as the Dice coefficient, Intersection over Union (IoU), pixel accuracy, precision, recall, and loss function. The results demonstrated that AdamW and RMSprop were more effective in improving segmentation accuracy, whereas methods like Adagrad and Adafactor exhibited lower performance. Additionally, the proposed approach was compared with previous studies conducted on similar datasets, achieving competitive or superior results. This study highlights the impact of optimization algorithms on model performance in lumbar spine segmentation and contributes to deep learning-based medical image analysis research.