A dual-decoder banded convolutional attention network for bone segmentation in ultrasound images

Abstract

Background: Ultrasound (US) has great potential for application in computer-assisted orthopedic surgery (CAOS) due to its non-radiative, cost-effective, and portable traits. However, bone segmentation from low-quality US images has been challenging. Traditional segmentation methods cannot achieve satisfactory results due to their high customization and dependence on bone morphology. Existing deep learning-based methods make it difficult to ensure efficient and accurate segmentation due to the ignorance of prior knowledge of bone features during feature learning.

Purpose: This paper aims to systematically investigate feature extraction and segmentation methodologies of bone US images and then proposes an innovative convolutional neural network to address the need for precise and efficient bone structure extraction in CAOS.

Methods: This paper has proposed a dual-decoder banded convolutional attention network (BCA-Net), which takes the raw US image as input and simplified U-Net as the baseline network. Multiscale banded convolution kernels are employed internally in the BCA-Net model, leveraging the prior knowledge that bone surfaces in US images are exhibited as bright bands of a few millimeters in width. Additionally, a shared encoder to extract input features and two independent decoders to generate outputs for the bone surface and bone shadow mask are integrated into the BCA-Net model, leveraging the prior knowledge that US bone surfaces manifest low-intensity hollow shadows below. Then, a new task consistency loss is introduced that can utilize inter-task dependency fully and enhance the performance of our model. In the network construction process, a dataset containing 1623 sets of US images was adopted, and a five-fold cross-validation strategy was divided into the training and validation sets for the model's training and validation. Many vital metrics were introduced to comprehensively evaluate the model performance, including overlap, edge distance, area under curve, and model efficiency. Finally, the model performance was subjected to a confidence interval, Tukey's honest significant difference, and Cohen's d statistics at a significance level (5%) to ensure the accuracy and reliability of the obtained findings.

Results: The experimental results show that the BCA-Net model performs well in the bone surface segmentation task. Its average Dice coefficient reaches 87.51%, 4.04% higher than U-Net's, proving its superior bone surface segmentation accuracy. Meanwhile, the average distance error is 0.2 mm, 0.33 mm lower than U-Net's, highlighting its accuracy in detail capture and boundary recognition. Using a confidence distance threshold of 1.02 mm, the Dice coefficient of the BCA-Net model exceeds 98%, an improvement of 1.87% over U-Net's, which is highly consistent with manual labeling. The BCA-Net model achieves a statistical significance of p-values < 0.05 in the above accuracy comparisons. In addition, the BCA-Net model has a small parameter count (5.58 M) and high computational efficiency (35.85 frames per second), further validating its excellent potential in bone surface segmentation tasks.

Conclusions: The proposed method achieves excellent performance with high accuracy and efficiency, aligning well with clinical requirements and holding excellent potential for advancing the utilization of US images in CAOS.

Keywords: banded convolution; bone segmentation; deep learning; dual‐decoder network; ultrasound images.