A Hybrid Efficient U-Net Framework for Detection of Anterior Belly of the Digastric Muscle on Ultrasonography

Abstract

The digastric muscle is an important muscle involved in functions such as chewing and swallowing. Ultrasonography is the preferred method for imaging the soft tissues of the head and neck but is highly operator-dependent. Artificial intelligence, particularly deep learning-based segmentation models, has the potential to improve the accuracy and precision of ultrasound images. In this study, a MultiResUNet-Fusion model including residual blocks, multiscale feature fusion, and SE blocks was developed for segmentation of the anterior belly of the digastric muscle. The model was trained on 198 ultrasound images from 99 participants. Combo Loss (a combination of Binary Cross-Entropy and Dice Loss) was used to train the model and segmentation metrics such as F1-score, Intersection over Union (IoU) and Dice Co-efficient were used to evaluate performance. The proposed MultiResUNet-Fusion model provided high accuracy and reliability for the segmentation of the anterior belly of the digastric muscle. The proposed MultiResUNet-Fusion model demonstrated high performance by achieving F1 score (95.38%) and IoU (91.17%). The visual results showed that the segmentation masks of the MultiResUNet-Fusion models provided predictions close to the real labels, and all models generally localized the region of interest accurately. The MultiResUNet-Fusion model provides high accuracy in low-contrast ultrasound images, making it suitable for clinical applications. The model can contribute to clinical diagnostic processes with its ability to accurately detect small and large structures. Future studies can increase the generalization capacity of the model by testing it in different modalities.