Enhancing fetal ultrasound image quality assessment with multi-scale fusion and clustering-based optimization

Fetal ultrasound image quality assessment (FUIQA) requires experienced physicians to identify anatomical structures and overall information within the image, providing a quality evaluation based on their subjective perception. Obtaining high-quality images can aid in clinical diagnosis and serve as reliable input for downstream tasks such as structure detection, organ segmentation, and disease diagnosis. However, the objective and accurate assessment fetal ultrasound image quality using deep learning poses several challenges. The FUIQA process involves an interaction between objective analysis and subjective judgment. Quality annotation requires careful attention to both global and local information, making it time-consuming and labor-intensive. Additionally, optimizing deep models with case-wise loss functions may result in suboptimal performance on global evaluation metrics compared to directly optimizing for global correlation consistency. In this paper, we construct a novel deep learning-based FUIQA framework to address these challenges. Our contributions are threefold. First, we propose a multiscale quality-aware fusion module that fuses and enhances feature representations at different scales, thereby ensuring effective feature perception and improving model performance. Second, we introduce a clustering-based mix-up (CMX) strategy to generate a sufficient number of rich pseudo-samples, alleviating the problem of insufficient training samples. Third, considering that the pseudo-samples generated by CMX can effectively simulate the global distribution under the batch-based training approach, we design a novel global correlation consistency loss to directly learn global evaluation metrics, ensuring consistency between training and testing objectives. Extensive experiments on five FUIQA datasets demonstrated that our framework outperforms other strong competitors.