FNBUI-NET: A multi-task model for fetal nasal bone ultrasound image defect detection and classification

Abstract

In prenatal ultrasound diagnosis, the accurate identification and timely assessment of the fetal nasal bone are crucial for monitoring the growth and development of the fetus’s face. However, the small size and complex anatomy of the fetal nasal bone make it difficult for doctors to determine if there are any defects. Motivated by this critical and intricate ultrasound task, a multi-task deep learning model, FNBUI Network (FNBUI-NET), was designed to detect fetal nasal bone defects and classify fetal nasal bone ultrasound images (FNBUI) in real-time. The model reconsiders and balances the relationship between inference speed and computational accuracy in object detection algorithms, innovatively proposing the integration of the ShuffleNetV2 (Shuffle_Block) network with the Content-Aware ReAssembly of Features (CARAFE) module to process data tensors of varying sizes from clinical settings, and employed Coordinate Attention (CA) for feature compression encoding to extract consistency constraints within chaotic feature distributions. The FNBUI dataset is also proposed to support the research on fetal nasal bone detection and classification, which 1104 ultrasound images with three objects (i.e., normal, dysplasia, absence) of fetal nasal bones from 731 mid-pregnancy women for training and testing the performance of FNBUI-NET. The experiments show that FNBUI-NET can accurately classify fetal ultrasound images and detect nasal bone defects, achieving an [email protected] of 0.985 in nasal bone detection and an accuracy of 0.981 in classification, outperforming 18 popular current methods. It provides a valuable tool for less experienced and younger sonologists, assisting them in the examination and diagnosis of fetal nasal bone defects.