1 Department of Medicine, Division of Physical Medicine and Rehabilitation, University of Toronto, Toronto, ON, Canada.
2 Physical Sciences, Sunnybrook Research Institute, Toronto, ON, Canada.
Ultrason Imaging. 2019 Jul;41(4):231-246. doi: 10.1177/0161734619839980. Epub 2019 Apr 16.
Research involving B-mode ultrasound imaging often requires user defined regions of interest (ROIs) for analysis, traditionally drawn/selected by a trained operator. This manual process is incredibly time consuming and subjective. Here, we propose a fast and simple method of detecting the average location of aponeurosis layers in ultrasound images of the upper trapezius to place a rectangular ROI for quantitative image analysis. A total of 56 B-mode ultrasound images were analyzed, where rectangular ROIs were manually placed in the skeletal muscle by two trained operators. Interoperator agreement was determined between the ROI border locations using intercorrelation coefficient (ICC). Next, our automatic algorithm was applied (image thresholding, binary masking, and pixel intensity peak detection), estimating the mean position of both aponeurosis layers for rectangular ROI placement. The automatic estimation method was compared with the manual (visual) method by various statistics ( t test, linear correlation, Bland-Altman plot). The performance was also evaluated under additive noise conditions (Speckle). Finally, agreement of the overlapping ROI area between the manual and automatic methods was also computed. Performance of the automatic method compared with manual placement was excellent for both the superficial and deep ROI borders, performing consistently even with additive noise (error <0.674 ± 1.69 mm). Manual measurements indicated excellent consensus (ICC = 0.902) between operators. The overlapping area between the manual and automatic measurements demonstrated good agreement (90.65 ± 11.3%). With constraints, our method is robust even under large levels of noise addition making the automatic algorithm an acceptable replacement for manual ROI placement in the upper trapezius.
研究涉及 B 模式超声成像时，通常需要用户定义感兴趣区域 (ROI) 进行分析，传统上由经过培训的操作人员进行绘制/选择。这个手动过程非常耗时且主观。在这里，我们提出了一种快速简单的方法来检测上斜方肌超声图像中腱膜层的平均位置，以便为定量图像分析放置矩形 ROI。总共分析了 56 张 B 模式超声图像，其中两个经过培训的操作人员在骨骼肌中手动放置了矩形 ROI。使用互相关系数 (ICC) 确定 ROI 边界位置的操作者间一致性。然后，我们应用了自动算法（图像阈值、二值掩模和像素强度峰检测），估计矩形 ROI 放置时两个腱膜层的平均位置。通过各种统计（ t 检验、线性相关性、Bland-Altman 图）比较自动估计方法与手动（视觉）方法。还计算了手动和自动方法之间重叠 ROI 区域的一致性。对于浅层和深层 ROI 边界，自动方法与手动放置相比表现出色，即使在添加噪声的情况下也能保持一致（误差 <0.674 ± 1.69 mm）。手动测量值显示出操作者之间的良好一致性（ICC = 0.902）。手动和自动测量之间的重叠区域显示出良好的一致性（90.65 ± 11.3%）。在有约束的情况下，即使添加了大量噪声，我们的方法也很稳健，这使得自动算法成为上斜方肌中手动 ROI 放置的可接受替代品。
