Deep learning for diagnosis of congenital heart disease in fetus using MRI and ultrasound

Lead Research Organisation: King's College London
Department Name: Imaging & Biomedical Engineering


Aim of the PhD Project:

The aim of this project is to design a deep learning approach for diagnosis of congenital heart disease using fetal MRI and ultrasound, which includes visualisation of cardiac anatomy from motion corrected MRI, alignment of ultrasound of moving heart and blood flow, and interpretable deep learning for prediction post-natal outcomes.

Project Description / Background:

Congenital Heart Disease (CHD) is the most common congenital malformation, affecting 8 out of a thousand births. Up to 25% of babies with CHD have a major abnormality [1] and delayed diagnosis is associated with increased mortality [2], [3]. The standard clinical modality for prenatal diagnosis is ultrasound, however in many cases the diagnosis may be incomplete, due to difficulties in visualising the structure and topology of the major vessels [4] caused by artefacts and variable ultrasound image quality. Recently we have shown that motion corrected fetal MRI [5] improved diagnostic quality in 90% of the cases [6]. However, the motion corrected fetal cardiac MRI has a number of disadvantages. Firstly, the MRI has to be reconstructed from stack of slices corrupted by interslice motion and major motion, common at around 20 weeks pregnancy when the exam is required, results in motion correction algorithm failures. For this reason, the exam is currently postponed until 30 weeks of pregnancy. Secondly, the current pipeline requires time-consuming manual input, such as manual reorientation and segmentation, that prevents the translation to routine clinical practice. Finally, even if motion-corrected 3D MRI is good quality, it has low spatial resolution that prevents visualisation of anatomical details such as valves and does not capture movement of the heart. Ultrasound, on the other hand, though poor for visualisation of 3D topology, offers higher spatial and temporal resolution as well as measurement of the blood flow in fetal heart and vessels. We hypothesise that combining anatomical and functional multimodal markers into the common reference frame will open possibilities enhanced and more accurate diagnosis and prediction of outcomes for fetuses with CHD.

The aim of this project is to design a deep learning approach for diagnosis of CHD at mid-pregnancy that combines advantages of ultrasound and MRI. The project will progress in four stages:

Development of deep learning-based motion correction algorithm that allows for fully automatic and reliable reconstruction of fetal MRI at around 20 weeks of pregnancy
Development of automatic segmentation and 3D visualisation of fetal heart in motion corrected MRI
Registration of fetal US data with the motion corrected MRI to facilitate qualitative diagnosis by joint assessment of fetal heart anatomy, motion and blood flow
Interpretable machine learning to discover new complex biomarkers of abnormalities that are currently difficult to diagnose prior to birth, such as coarctation of the aorta, and prediction of post-natal outcomes in these babies.


10 25 50

Studentship Projects

Project Reference Relationship Related To Start End Student Name
EP/S022104/1 01/10/2019 31/03/2028
2442175 Studentship EP/S022104/1 01/10/2020 30/09/2024 Paula Ramirez Gilliland