Deep learning approaches enable automatic interpretation for fetal screening, which yields useful clinical information for prenatal medical care. However, utilising deep learning for fetal screening analysis is still challenging. Shadow artefacts in fetal ultrasound imaging may conceal anatomical structures, and thus result in poor anatomy visualisation and inaccurate image interpretation. The utilisation of learning-based fetal screening analysis algorithms in patient care at scale is hindered by the data difference of images obtained from different acquisition devices, hospitals and geographic regions. Shortage of workforce and requirement of special expertise lead to insufficient and coarse prior knowledge for fetal screening analysis. This thesis aims at developing deep learning methods for fetal screening analysis with little or no supervision, specifically on shadow artefacts and anatomical classification across different datasets.

We first propose learning-based methods to estimate shadow confidence maps for fetal ultrasound images based on coarse and weak image annotations. The predicted dense shadow confidence maps show the probability of each pixel being shadow regions and can provide extra information for improving the performance of downstream automatic image analysis algorithms such as fetal ultrasound classification, multi-view image fusion and biometric measurement. We then address the problem of the constrained utilisation of fetal ultrasound standard plane classification model across different datasets. A deep learning-based method is proposed to align anatomical features between different datasets and it enables the consistent performance of fetal ultrasound standard plane classification on images acquired from different devices. Finally, we investigate the generalisation of task-specific models when images from different clinics do not share the same anatomical categories. We propose learning-based methods to separate the anatomical features from all other types of features, and thus to learn generalisable anatomical features. The proposed methods enable the classification of unseen anatomical categories, which can help clinicians from different clinical sites in a wide range of geographic areas to use the same fetal ultrasound standard plane classification model for the analysis of their own data.