Effects of convolutional neural network models on the segmentation of the interscalene brachial plexus in ultrasound imaging for radiomics.
Authors
Affiliations (2)
Affiliations (2)
- Department of Anesthesiology, Beijing Jishuitan Hospital, Capital Medical University, Beijing, China.
- Department of Pain Management, Beijing Jishuitan Hospital, Capital Medical University, Beijing, China.
Abstract
The application of artificial intelligence (AI) in ultrasound (US)-guided regional anesthesia has expanded, particularly in enhancing the accuracy, safety, and training effectiveness of procedures through deep learning-based anatomical segmentation. This study aimed to develop and validate an automatic segmentation model for supraclavicular-to-interscalene brachial plexus block (ISB) using the You Only Look At CoefficienTs (YOLACT) algorithm and to compare its performance with that of a U-Net model. A total of 1,100 patients scheduled to undergo ISB were enrolled. US images encompassing the anatomical range from the supraclavicular fossa to the C7 vertebral level were acquired by experienced anesthesiologists. Images were annotated to identify the brachial plexus nerve, anterior scalene muscle (ASM), middle scalene muscle (MSM), and subclavian artery (SA). YOLACT and U-Net models were trained for automatic segmentation. Model performance was assessed using Intersection over Union (IoU), Dice similarity coefficient (DSC), Hausdorff distance (HD), the proportion of images with a brachial plexus nerve IoU > 0.5, and segmentation accuracy metrics. A total of 6,600 US images were analyzed. The YOLACT model demonstrated significantly higher IoU and DSC values and lower HD values compared with the U-Net model for segmentation of the brachial plexus nerve, ASM, MSM, and SA (P<0.001). The proportion of images with a brachial plexus nerve IoU greater than 0.5 was also significantly higher with YOLACT (P<0.001). An automatic segmentation model for ISB, spanning the supraclavicular region to the C7 level, was developed using the YOLACT algorithm. Although quantitative performance metrics favored YOLACT over U-Net, subjective accuracy assessments were comparable between models. Further studies using larger datasets are required to clarify the potential clinical applicability of this approach.