Yu Tianxiang, Tsui Po-Hsiang, Leonov Denis, Wu Shuicai, Bin Guangyu, Zhou Zhuhuang
Department of Biomedical Engineering, College of Chemistry and Life Science, Beijing University of Technology, Beijing 100124, China.
Department of Medical Imaging and Radiological Sciences, College of Medicine, Chang Gung University, Taoyuan 333323, Taiwan.
Sensors (Basel). 2024 Nov 25;24(23):7510. doi: 10.3390/s24237510.
The detection of fetal ultrasound standard planes (FUSPs) is important for the diagnosis of fetal malformation and the prevention of perinatal death. As a promising deep-learning technique in FUSP detection, SonoNet's network parameters have a large size. In this paper, we introduced a light pyramid convolution (LPC) block into SonoNet and proposed LPC-SonoNet with reduced network parameters for FUSP detection. The LPC block used pyramid convolution architecture inspired by SimSPPF from YOLOv6 and was able to extract features from various scales with a small parameter size. Using SonoNet64 as the backbone, the proposed network removed one of the convolutional blocks in SonoNet64 and replaced the others with LPC blocks. The proposed LPC-SonoNet model was trained and tested on a publicly available dataset with 12,400 ultrasound images. The dataset with six categories was further divided into nine categories. The images were randomly divided into a training set, a validation set, and a test set in a ratio of 8:1:1. Data augmentation was conducted on the training set to address the data imbalance issue. In the classification of six categories and nine categories, LPC-SonoNet obtained the accuracy of 97.0% and 91.9% on the test set, respectively, slightly higher than the accuracy of 96.60% and 91.70% by SonoNet64. Compared with SonoNet64 with 14.9 million parameters, LPC-SonoNet had a much smaller parameter size (4.3 million). This study pioneered the deep-learning classification of nine categories of FUSPs. The proposed LPC-SonoNet may be used as a lightweight network for FUSP detection.
胎儿超声标准平面（FUSP）的检测对于胎儿畸形的诊断和围产期死亡的预防至关重要。作为FUSP检测中一种有前景的深度学习技术，SonoNet的网络参数规模较大。在本文中，我们将轻量级金字塔卷积（LPC）模块引入SonoNet，并提出了用于FUSP检测的网络参数减少的LPC - SonoNet。LPC模块采用了受YOLOv6的SimSPPF启发的金字塔卷积架构，能够以较小的参数规模从不同尺度提取特征。以SonoNet64作为主干网络，所提出的网络移除了SonoNet64中的一个卷积模块，并用LPC模块替换了其他模块。所提出的LPC - SonoNet模型在一个包含12400张超声图像的公开可用数据集上进行训练和测试。这个包含六个类别的数据集进一步被划分为九个类别。图像按照8:1:1的比例随机分为训练集、验证集和测试集。对训练集进行数据增强以解决数据不平衡问题。在六类别和九类别分类中，LPC - SonoNet在测试集上分别获得了97.0%和91.9%的准确率，略高于SonoNet64的96.60%和91.70%的准确率。与具有1490万个参数的SonoNet64相比，LPC - SonoNet的参数规模要小得多（430万个）。本研究开创了九类FUSP的深度学习分类。所提出的LPC - SonoNet可作为用于FUSP检测的轻量级网络。
