Singular pooling: a spectral pooling paradigm for second-trimester prenatal level II ultrasound standard fetal plane identification

Ultrasound examination has become a vital mid-term prenatal screening method due to its numerous benefits. However, identifying fetal ultrasound sections poses significant challenges and requires the expertise of experienced doctors, especially when dealing with unfavorable fetal positions and image artifacts. The shortage of skilled doctors and the complexity of these screenings highlight the necessity of artificial intelligence to support medical professionals. Most current research targets a few specific fetal ultrasound sections for particular tasks. In reality, mid-pregnancy level II ultrasound assessments involve around ten different fetal sections. Therefore, there is a critical need for algorithms that can recognize these level II fetal sections. Furthermore, the standard ultrasound planes located in the same region exhibit a high degree of similarity. The traditional fetal slice algorithm such as ResNet generates feature maps that are sparse matrices. As a result, employing global average pooling (GAP) layers can lead to significant information loss, making it challenging to effectively differentiate between similar views. In this study, we introduce a deep-learning network designed for second-trimester anatomy ultrasound standard plane recognition. Our approach incorporates a novel spectral pooling paradigm. We introduce singular pooling to enhance information extraction from feature maps and integrate this singular pooling layer into ResNet. This methodology is tested on our dataset, which includes ten distinct standard fetal ultrasound sections from second-trimester level II examinations. Our model achieves exceptional performance, with an accuracy of 92.07% and an F1-Score of 0.919. The implementation of singular pooling significantly improves the original ResNet model. This method can effectively aid in the identification of standard plane recognition in second-trimester anatomy ultrasounds. Furthermore, we validated the efficacy of our model by applying it to two additional datasets, thereby further demonstrating its efficiency and applicability. Our code is available at https://github.com/sysll/vectore/tree/ master.