Anatomical structures detection using topological constraint knowledge in fetal ultrasound

The accurate recognition of anatomical structures in fetal ultrasound images is crucial for prenatal diagnosis and determining ultrasound standard planes. However, this task can be arduous and time-consuming for sonographers. Existing methods usually treat anatomical structures as isolated individuals, ignore their interconnections, and lead to false and missed detections due to the similarities between these structures. In this paper, we propose a key anatomical structure-aware module, named KASM, which consists of class-aware representation learning and topological information interaction and is designed to integrate seamlessly into any proposal feature-based detection framework. Given the potential issue of information interactions between proposal features leading to feature over-smoothing, our approach adopts class-aware representation learning as an initial step. This process aims to enhance the intra-class compactness of proposal features while simultaneously dispersing inter-class features. The objective is to derive highly discriminative proposal features, thus facilitating subsequent accurate information interactions. Subsequently, proposal features are modeled using topological constraint relationships between anatomical structures and informative interactions are performed for structure recognition. We conduct extensive experiments on collected fetal ultrasound datasets in femoral, abdominal, and thalamic planes. The results consistently demonstrate that KASM effectively mitigates issues such as false detection and missed detection of anatomical structures across different baseline models, e.g., integrating KASM in Faster R-CNN reduces the false detection rate from 49.7% to 38.4% (−11.3%) and the missed detection rate from 60.8% to 57.0% (−3.8%) under the confidence threshold of 0.2. This underscores the broad applicability and efficacy of KASM in addressing these challenges. © 2024