Doctor of Engineering (EngD) – Shortcut mitigation for medical imaging foundational models for segmentation (Canon Medical Research Europe and University of Edinburgh)

Project Description

Imaging foundational models (IFM) are often defined as versatile deep learning models performing a task-agnostic feature extraction using a robust backbone pre-trained via self-supervised, semi-supervised or fully supervised methods on a large-scale, diverse dataset. They are often initially developed for a wide range of clinical use-cases or downstream tasks. Once this initial training is done, they can be adapted to different downstream tasks through an adaptation phase. This adaptation phase can take multiple forms from prompting to extensive fine-tuning. While the original training phase requires huge datasets and significant computational resources, the adaptation phase is often less data and computation hungry. Interestingly, while the aim of IFM is to provide universal imaging models, many recent studies have highlighted both their struggle to generalize and their tendency to encode bias. This project aims to investigate shortcut mitigation techniques in the adaptation phase of imaging foundational models. The goal is to propose methods that would mitigate existing shortcuts without requiring identification and labeling of said shortcuts. While this project is aimed at improving the usability of imaging foundational models for segmentation, the proposed methods will be applicable to any deep learning segmentation method requiring adaptation such as fine-tuning. 

The researcher will be based within the AI CoE research team at Canon Medical Research Europe, in Edinburgh. Canon Medical are one of the largest manufacturers of medical imaging equipment, including X-ray, CT, MRI, nuclear medicine, PET and ultrasound imaging. The student will be able to benefit from the company’s flexible working policy and work both on-site and remotely. 

Please find the company’s public EDI policy on our website: https://research.eu.medical.canon/careers/diversity-inclusion/

CDT Essential Criteria

A Masters level degree (MEng, MPhys, MSc) at 2.1 or equivalent.

Desire to work collegiately, be involved in outreach, undertake taught and professional skills study.

Project Essential Criteria

Experience in machine learning/deep learning

Experience in image processing/computer vision

Project Desirable Criteria

An interest in healthcare/medical imaging

The CDT

The CDT in Applied Photonics provides a supportive, collaborative environment which values inclusivity and is committed to creating and sustaining a positive and supportive environment for all our applicants, students, and staff. For further information, please see our ED&I statement: https://bit.ly/3gXrcwg.

Forming a supportive cohort is an important part of the programme and our students take part in various professional skills workshops, including Responsible Research and Innovation, and attend outreach training.