The project focuses on development of automated deep learning solution for measurement, detection and characterisation of anomalies in fetal brain MRI with an emphasis on integration into radiological workflow and clinical translation.
(a) The scientific basis:
Fetal MRI is increasingly used to assess suspected brain abnormalities, offering greater anatomical detail than ultrasound. Yet, current reporting depends on subjective visual assessment and manual 2D measurements, limiting accuracy and consistency. Volumetric analysis is rarely performed due to the impracticality of manual segmentation. This project aims to develop deep learning tools for automated 3D segmentation and anomaly detection in pathological fetal brain MRI, enabling objective, reproducible reporting. Building on a validated pipeline (https://github.com/SVRTK/auto-proc-svrtk) and a large dataset (>3,500 scans) from the Department of Early Life Imaging and St.Thomas’ Hospital, the project has strong translational potential for clinical use in the NHS.
(b) The techniques and skills the student will develop:
The student will gain expertise in deep learning for medical imaging, focusing on 3D segmentation and anomaly detection using the MONAI framework (https://monai.io). They will learn to handle real-world imaging challenges such as anatomical variability, acquisition differences, and artefacts. Technical skills will include deep learning network training, anomaly detection, uncertainty quantification, and quality control pipelines and Docker containerisation. Clinically, the student will acquire knowledge of fetal neurodevelopment, MRI diagnosis, and regulatory frameworks for clinical AI software deployment.
(c) Aims of the project:
1. Develop and validate a deep learning pipeline for multi-regional segmentation of abnormal fetal brain anatomy.
2. Design and develop an anomaly detection framework tailored for pathological fetal brain MRI.
3. Integrate both tools into a prototype radiology reporting pipeline, optimised for clinical translation.
(d) Objectives:
Year 1: Literature review and technical evaluation of existing segmentation and anomaly detection networks; hands-on training using synthetic and normal fetal MRI data; submission of a review publication.
Year 2: Implement and validate 3D segmentation pipeline for the most common brain anomalies; submit first-author publication.
Year 3: Develop and evaluate anomaly detection algorithms; publish findings in peer-reviewed journal.
Year 4: Integrate pipelines into deployable software; perform validation and robustness testing; submit publication and prepare for clinical evaluation.
(e) A summary of a potential 3-month rotation project:
Train and evaluate a 3D segmentation model for volumetric analysis in cases of ventriculomegaly—the most common fetal brain MRI anomaly. The student will gain practical experience with neural network training, evaluation metrics, and data curation.