Towards Explainable and Generative AI for Fetal Ultrasound Image Analysis

Artificial intelligence (AI) has transformed medical imaging and is becoming the standard for fetal ultrasound image preprocessing and analysis. However, responsible clinical deployment remains challenging due to concerns about transparency, bias, and robustness. This thesis adopts a human-centered perspective on explainable AI as a practical toolkit that can help people understand the strengths and limitations of AI models and datasets, supporting the responsible deployment of robust and fair AI systems. Motivated by the power of counterfactual explanations that explore realistic “what-if” scenarios and bridge explainable and generative AI, we focus on developing novel methods for counterfactual generation that serve both model-centric and data-centric analysis to understand model decisions and reveal biases in the data. Building on this perspective, we leverage modern generative diffusion models to build practical tools for unsupervised fetal brain anomaly detection and counterfactual generation in fetal image quality assessment, enabling downstream applications such as diagnosis of rare fetal brain cases and personalised training for novice sonographers. In addition, we advance the state-of-the-art in predicting spontaneous preterm deliveries from transvaginal ultrasound while enhancing explainability through actionable, clinician-oriented AI feedback. We hope our contributions outline a practical path toward establishing explainable and generative AI in fetal ultrasound image analysis and prenatal care.