Original ultrasound images and Hough transformed images

The morphological characteristics of skeletal muscles, such as fascicle orientation, fascicle length, and muscle thickness, contain valuable mechanical information that aids in understanding muscle contractility and excitation due to commands from the central nervous system. Ultrasound (US) imaging, a non-invasive measurement technique, has been employed in clinical research to provide visualized images that capture morphological characteristics. However, accurately and efficiently detecting the fascicle in US images is challenging. In the current study, we employed computer vision techniques based on shallow/light neural networks (SNN) to detect the locations and orientations of fascicles in US images. To enhance the linear/tubular feature of the fascicle, we developed a weighted Hough transform algorithm to transform the original gray-scaled images to a weighted Hough space (WHS), which is expected to reduce the demand in layers for the neural network (NN) design. Compared to many baseline methods, including Vgg, ResNet, AlexNet, Unet, and UltraTrack, the proposed SNN was found to be more accurate in detection. Among the SNN methods, a single-layer convolutional neural network outperformed the others. Our study found that WHS usually improved the detection accuracy for SNN models, and the regression with an $\mathcal{L}_{2}$ regularization provided satisfactory detection even without WHS, which is suitable for real-time applications. Moreover, our proposed methods are robust to disturbances like data loss, time delay, and out-of-order, as they do not require the image frames to be closely interlinked or temporally related.  The path of the file is stated in the code, which can be found here: https://github.com/XBao06093030/SCNN_for_Fascicle_Detection/blob/main/SCNN_sample.ipynb

骨骼肌的形态学特征，如肌束走向、肌束长度与肌肉厚度，蕴含着宝贵的力学信息，有助于理解中枢神经系统指令介导的肌肉收缩与兴奋过程。超声成像（Ultrasound, US）作为一种无创检测技术，已被应用于临床研究，以获取能够捕捉肌肉形态特征的可视化图像。然而，在超声图像中精准且高效地检测肌束仍颇具挑战。本研究采用基于浅层/轻量神经网络（Shallow/Light Neural Networks, SNN）的计算机视觉技术，以检测超声图像中肌束的位置与走向。为强化肌束的线性/管状特征，我们开发了一种加权霍夫变换算法，将原始灰度图像转换至加权霍夫空间（Weighted Hough Space, WHS），此举有望降低神经网络（Neural Network, NN）设计对网络层数的需求。与包括VGG、ResNet、AlexNet、U-Net以及UltraTrack在内的多种基线方法相比，所提出的浅层/轻量神经网络在检测精度上更具优势。在各类浅层/轻量神经网络方法中，单卷积层神经网络的表现最优。本研究发现，加权霍夫空间通常可提升浅层/轻量神经网络模型的检测精度；而采用ℓ₂正则化的回归模型即便不借助加权霍夫空间，也能获得令人满意的检测效果，适用于实时应用场景。此外，所提出的方法对数据丢失、时延、时序混乱等干扰具有鲁棒性，因为其无需图像帧之间紧密关联或存在时序相关性。代码路径已在代码中说明，可通过以下链接查看：https://github.com/XBao06093030/SCNN_for_Fascicle_Detection/blob/main/SCNN_sample.ipynb