Semi-Supervised Learning for Foetal Anomaly Detection Using Limited Annotated Ultrasound

Foetal anomaly detection is a critical task in prenatal care, requiring early diagnosis for appropriate medical intervention. However, the scarcity of annotated ultrasound data limits the effectiveness of supervised deep learning models. In this work, the authors made use of the semi-supervised learning (SSL) approach using the FixMatch algorithm, which uses both labeled and unlabelled data. The model applies weak and strong augmentations with confidencebased pseudo-labelling to refine its predictions. Experiments on a foetal ultrasound dataset demonstrate a significant improvement in anomaly detection accuracy. Compared to a supervised baseline, the FixMatch based model achieved a 7.1 % accuracy gain using only 20 % labeled data. Visualisations of ROC plots and performance matrices validate the efficacy of the proposed method. This study highlights the potential of SSL in medical image analysis, reducing dependency on annotated datasets.