University of Cambridge Home Department of Engineering

F-AHG-2: Machine learning for segmentation of cochlear CT scans

A previous 4th year project developed an effective method for segmenting the human cochlea in clinical CT scans. That project, and indeed this one, was motivated by cochlear implants. The precise positioning of the implant within the cochlea can have a profound effect on the hearing outcome. And yet, when planning implantation surgery, it is difficult for the surgeon to take the individual's particular cochlear size and shape into account, given the low resolution of clinical CT images, and the difficulty in segmenting the cochlea from the surrounding temporal bone.

While the previous project took a traditional, model-fitting approach to the segmentation task, other researchers have attempted machine learning approaches. See, for example, the work of Heutink and Neves. The aim of this project is to implement a machine learning approach, and compare its performance with the traditional model-fitting approach.

An important point to note is that this project's supervisor is not a machine learning expert. So students should not apply if they would require supervision on this aspect of the project. But the project may appeal to a competent machine learning practitioner who would like to apply their expertise to an interesting medical problem. Supervision will be available on all other aspects of the project, including provision and preparation of training and testing data, and evaluation against the model-fitting approach.

This project is offered in collaboration with Professor Manohar Bance and Dr Chloe Swords at Addenbrooke's Hospital. The project would suit somebody who can read the Heutink and Neves papers, and would know how to re-implement their work. Having taken Module 3G4 and Project GG2 would also be an advantage.

Model-based segmentation of the human cochlea and semicircular canals in clinical CT images. The aim of this project is to investigate alternative, machine-learning methods for segmentation.
© Cambridge University Engineering Department
Information provided by Andrew Gee
Last updated: April 2023