Skip to main content
Menu

Research Studentship in Deep Learning for Medical Imaging

Research Studentship in Deep Learning for Medical Imaging

Project: Improving the Reliability of Neural Networks for Medical Imaging

3.5-year D.Phil. studentship, fully funded for UK applicants, starting October 2022

Supervisors: Dr Konstantinos Kamnitsas

We are seeking a highly motivated DPhil student to study and improve the reliability of deep neural networks for medical image analysis.

Advances in Machine Learning (ML) have revolutionized computer vision and image analysis. This could tremendously benefit healthcare, where ML-powered tools could facilitate analysis of medical images and consequently diagnosis and treatment. The state-of-the-art methods, Deep Neural Networks (DNNs), are capable of accurate predictions “on average”. This is insufficient, however, for safe adoption in high-risk applications such as healthcare, where even a single wrong prediction can be disastrous. Moreover, research has shown that predictions of ML models can be influenced unexpectedly by characteristics of the input data. For example, ML models developed for treatment planning were found to achieve lower performance when processing data of persons from certain ethnic groups. Consequently, for safe integration of ML tools in high-risk applications, it is desired that ML methods provide mechanisms for quantifying the reliability of their output and explain the factors that contribute to the prediction. This would enable users to identify wrong predictions and excluding them from follow-up decision making. Current DNNs do not provide such mechanisms. This direly limits their large-scale adoption in healthcare and other high-risk applications.

Goal of this project is to develop ML methods for automatically detecting failures of DNN predictions, identifying what characteristics of inputs cause such failures, and develop frameworks for training models that are more robust to such characteristics. The approach taken will depend on the candidate’s skills and interests and can evolve during the course of this project, with guidance and support by the supervisor. Topics include but are not limited to:

  • Causality, uncertainty estimation and out-of-distribution detection approaches for developing mechanisms that can detect faulty model predictions and the data characteristics that cause them.
  • Generative modelling, domain adaptation and self-supervised learning for learning more reliable and generalizing models from unlabelled data.
  • Distributed federated learning, which can enable the learning from of decentralized data held at different collaborating institutions. This can allow analysis of method reliability at larger scale and enable training more robust, better generalizing models.

The project ultimately aims to catalyse safe adoption of ML for medical image analysis. At the same time, the investigated research questions are fundamental in ML research and of relevance to many other high-risk applications, such as autonomous driving, amplifying the impact of this project.

The successful applicant will be a member of the Biomedical Image Analysis cluster (BioMedIA website) that is based in the Institute for Biomedical Engineering (IBME website). The BioMedIA cluster is a world-leading group in machine learning for medical image analysis. The applicant will have access to computing resources of IBME and the Big Data Institute (BDI website), as well as access to unique, real-world clinical databases. We are committed to equality, diversity and inclusion, and we are keen to attract applicants from a wide variety of backgrounds.

Eligibility

This studentship is funded through the UK Engineering and Physical Sciences Research Council (EPSRC) Doctoral Training Partnership and is open to Home students (full award – home fees plus stipend). Full details of the eligibility requirements can be found on the UK Research and Innovation website (see here for EU and international guidance).

There is very limited flexibility to support international students. If you are an international student and want to apply for this studentship, please contact the supervisor to see whether the flexibility might be available for you.

Award Value

This studentship is open to Home students only. It will cover the course fees as well as a stipend for 3.5 years. The stipend (tax-free maintenance grant) is at least £15,609 p.a. for the first year, and at least this amount for a further two and a half years.

Candidate Requirements

Prospective candidates will be judged according to how well they meet the following criteria:

  • A first class (or strong 2:1) honours degree or Distinction Masters level degree in Engineering, Computing, Mathematics, Physics, or relevant discipline.
  • Strong Mathematical and Analytical skills
  • Strong Programming skills (at least in Python)
  • Knowledge of basic Machine Learning principles (e.g. from undergraduate courses, projects or work experience)
  • Excellent written and spoken communication skills in English

The following skills are also desirable but not essential:

  • Previous experience with processing images. Experience analyzing other types of data is also of beneficial.
  • Previous experience with Software Engineering (e.g. demonstrated with contributions to Open Source projects or related work experience)

Application Procedure

Informal enquiries are encouraged and should be addressed to Dr. Konstantinos Kamnitsas (kamnitsas.k@gmail.com), incoming Associate Professor in Biomedical Imaging.

Candidates must submit a graduate application form and are expected to meet the graduate admissions criteria.  Details are available on the course page of the University website.

Please quote 22ENGBI_KK in all correspondence and in your graduate application.

Application deadline:  noon on 3 December 2021 

Start date: October 2022