CMSCNet: a context based lightweight musculoskeletal ultrasound image segmentation method

Abstract

Background: Musculoskeletal ultrasound (MSUS) is a non-invasive and non-intrusive method for examining muscles and bones. Therefore, MSUS image analysis plays a crucial role in the assessment and early detection of musculoskeletal disorders However, due to the complexity of noise in MSUS images, analyzing and interpreting these images is a tedious and time-consuming process. Currently, ultrasound devices still rely on manual methods to analyze structural parameters such as muscle thickness, penniform angle, and fascicle length. While recent advancements in deep learning have shown promise in automating image segmentation tasks, existing methods often require extensive computational resources and may not be suitable for real-time applications. There remains a need for lightweight and efficient models that can achieve high accuracy while reducing computational load. This study aims to address this gap by proposing a context-based lightweight deep-learning framework for the automatic segmentation of MSUS images.

Methods: We propose a context-based lightweight deep learning framework for the automatic segmentation of MSUS images, aiming to make the segmented images as close as possible to those manually annotated by professional doctors. This method is based on the U-Net architecture, with multi-layer perception modules added to the encoder and decoder to reduce the number of parameters and improve computational efficiency. Additionally, a dense atrous convolution module is used to extract contextual features from the images, improving segmentation accuracy, and a restructured convolution module for spatial and channel dimensions is employed to reduce the extraction of redundant features during the segmentation task. We applied our method to a publicly available leg muscle dataset to extract and analyze the morphological features of the penniform muscle and conducted ablation experiments.

Results: The results showed that the accuracy of our algorithm was 98.7%, the same as the U-Net architecture, but with only one-tenth of the parameters. The average intersection over union (IoU) also reached 0.7227. Our network can capture more details of the muscle aponeurosis and effectively focuses on the junctions between the aponeurosis and muscle fibers, showing minimal differences from the ground truth and achieving very high accuracy.

Conclusions: This method can automatically, quickly, and accurately extract the morphological features of penniform muscles, providing a basis for improving the accuracy of muscle pathology assessment, the precision of interventional therapy, and the scientific rigor of rehabilitation treatment.

Keywords: Musculoskeletal ultrasound (MSUS); U-Net; leg muscle; segmentation.

Figure 1

The framework of our proposed method. BN, batch normalization; DAC, dense atrous convolution; GeLU, Gaussian error linear unit; MLP, multilayer perceptron; ReLU, rectified linear unit; SCConv, spatial and channel reconstruction convolution module.

Figure 2

Structure of dense atrous convolution module. C, channel; Conv, convolution; H, height; W, width.

Figure 3

Structure of spatial and channel reconstruction convolution module. (A) The spatial reconstruction unit. (B) The channel reconstruction unit. Conv, convolution; GWC, group-wise convolution; PWC, point-wise convolution.

Figure 4

Result of data augmentation. (A) Original image. (B) Random horizontal flip. (C) Random vertical rotation. (D) Random rotation.

Figure 5

Result of CutMix. (A) Original image A. (B) Original image B. (C) CutMix.

Figure 6

Structure diagram of the ablation study downsampling. Conv, convolution; CRU, channel reconstruction unit; DAC, dense atrous convolution; MLP, multilayer perceptron; SRU, spatial reconstruction unit.

Figure 7

Visualization results of the ablation study.

Figure 8

Comparison of our network with other state-of-art algorithms.

Figure 9

Comparison of inference and training speed. GPU, graphics processing unit.

Figure 10

The segmentation performance on the public dataset is poor. (A) Poor-quality image with ambiguous boundaries and artifacts affecting fascia continuity; (B) misclassification of muscle fibers due to similar foreground-background grayscale values.