A research team from the public Macao Polytechnic University (MPU) has developed a generative artificial intelligence (AI) model capable of producing controllable foetal ultrasound images across multiple imaging planes, a breakthrough that could help address longstanding data shortages in prenatal medicine and AI-assisted diagnosis, according to an MPU statement yesterday.
The research was led by Tan Tao and Patrick Pang Cheong Iao, both associate professors from the university’s Faculty of Applied Sciences, alongside doctoral student Duan Yaofei, the statement said.
The project was conducted in collaboration with Shenzhen University, Shenzhen Duying Medical Technology Co. Ltd. and seven medical centres. Named “FetalFlex”, the model uses anatomical structures and multimodal information to generate foetal ultrasound images across different scanning planes, the statement noted, adding that it is the first unified framework capable of flexibly generating multiple foetal ultrasound views without requiring retraining for each specific plane. ‘
Foetal ultrasound examinations play a critical role in prenatal care, allowing doctors to assess foetal development and detect congenital abnormalities through images obtained from multiple angles, the statement said, adding that building comprehensive datasets for training radiologists and AI systems remains challenging, particularly for rare and complex foetal abnormalities that occur infrequently and have numerous subtypes.
According to the research team, FetalFlex addresses this challenge by generating both normal and abnormal foetal ultrasound images, including clinically meaningful abnormal cases even when no abnormal sample data are available, the statement said, noting that this capability fills an important gap in medical image generation and could support the development of more reliable diagnostic AI systems.
Extensive testing on multi-centre datasets showed the model achieved advanced image quality and strong agreement with radiologists’ visual assessments, the statement underlined, pointing out additional evaluations suggested that the generated abnormal images have potential clinical value.
The statement also said that incorporating FetalFlex-generated images into downstream deep learning tasks significantly improved model performance, aiming to offer a practical solution to the chronic shortage of foetal ultrasound data.
The research, titled “FetalFlex: Anatomy-Guided Diffusion Model for Flexible Control on Foetal Ultrasound Image Synthesis”, was published in the Medical Image Analysis journal, the statement said. The research team said that the technology could provide radiologists with more diverse training materials, including rare abnormal cases, while enabling doctors to simulate various conditions through anatomy-level image editing for educational and clinical purposes, the statement underlined.
This image provided by the public Macao Polytechnic University (MPU) yesterday shows the cover of the Medical Image Analysis journal.
