معصومه حسينقلي زاده وحيد

In this study, with the aim of automatically measuring the length of the calf muscle group from ultrasound images, the performance of two advanced segmentation architectures, including the transformer-based SegFormer model an‎d the convolutional network equipped with the Attention mechanism, Attention U-Net, was eva‎luated an‎d compared on synthetic an‎d real datasets. Measuring muscle length is a key indicator in biomechanical analysis, rehabilitation monitoring, diagnosing neurological an‎d muscular disorders, an‎d assessing injury risk. Therefore, achieving accurate segmentation of muscle fibers is a fundamental step in automating this process. The synthetic dataset included 2550 training frames, 450 validation frames, an‎d 30 test frames, while the real dataset contained 309 frames. The models were trained using the Binary Cross Entropy an‎d Focal Dice loss functions for SegFormer an‎d Attention U-Net, respectively, to investigate the impact of each loss function on metrics such as accuracy, recall, F1 score, an‎d IoU. The results showed that both models performed well in extracting muscle structures accurately. SegFormer, leveraging the transformer architecture an‎d lightweight decoder, demonstrated high capability in preserving image details an‎d offered fast training speeds. On the other han‎d, Attention U-Net, through its attention mechanism, showed higher sensitivity in identifying narrow boundaries an‎d smaller regions. These findings could pave the way for the development of reliable deep learning-based systems for automatic muscle length measurement in both research an‎d clinical settings.

بخش‌بندي تصوير،اندازه گيري طول ماهيچه، سونوگرافي، SegFormer، Attention U-Net، توابع خطا، امتيازF1، IoU

Image segmentation, muscle length measurement, ultrasound, SegFormer, Attention U-Net, loss functions, F1 score, IoU

Masoumeh Hosseingholizadeh vahid

Dr. Hamid Behnam