Trustworthy Machine Learning for Thyroid Image Analysis and Diagnosis

Half of women and one-third of men will develop a thyroid nodule at some point in their lives. 5-15% of these will be due to cancer. Ultrasound imaging provides first-line investigation in thyroid nodule diagnosis. If a nodule is deemed as indeterminate by radiologists, fine needle aspiration biopsy needs to be performed. A key challenge in current clinical workflow is that the interpretation of thyroid nodules in ultrasound images is labour intensive, time consuming and with inter-observer variability among radiologists. The aim of this project is to develop trustworthy machine learning algorithms that can efficiently interpret thyroid ultrasound images and enable faster diagnosis, algorithms for detection and classification of thyroid cancer in ultrasound imaging. The algorithms will provide explainable features for classification that can be understood by stakeholders concerned, who are the clinicians in this case. The project will leverage a large dataset accrued from experienced thyroid ultrasound radiologists of more than 20 years' experience at Hammersmith Hospital, which is the highest-volume thyroid surgical unit in the UK.