Aim: To develop a tool that identifies critical abnormalities on MRIbrain scans using deep learning.
A clinically validated tool that identifies brain tumours during brain CT or MRI scans does not exist. Immediate identification of a brain tumour potentially allows early intervention to improve clinical outcomes.Given the huge demand in scan referrals and given that only 1.4% of suspected patients referred by their GP for scanning will have a brain tumour, initial stratification is also essential. Clinical features canbe used in a prediction model.
Skills learnt:Machine learning, MRI translational design, statistics. Also broader topics includingCareers & Employability, Communication & Impact, Personal Effectiveness,Writing & Publishing. Complemented by faculty and departmental lectures;seminars; one-to-one supervisionsto develop skill in data handlingandanalysis.
Objectives:Year 1-2Build a classification model to stratify patients with undiscovered brain tumours for brain scan imaging(e.g. with headache). Implement several approaches andcontrast them,including:recurrent neural network (long short-term memory), Transformer and variational auto-encoder approaches.Training and hold-out testing for patient groups. Build a computer vision classification model to determine brain tumours from brain scans. Year 3Clean datasets for patient groups where symptom key words may improve classification accuracy.O4Validation of the classification model using routine clinical and imaging data. Complete analytical validation of classification model identifying relevant abnormalities onbrain MRI scans to demonstrate >0.95 accuracy/precision/recall/F1-score.Year 4Givea personalized output regarding the likelihood (including confidence intervals) and location of a brain tumour.