Artificial Intelligence with Human In The Loop for Automated Medical Image Contouring in Precision Oncology

Positron Emission Tomography (PET) yields quantitative images of regional in-vivo biology and biochemistry in the form of radioactive uptake, whilst Computed Tomography (CT) provides detailed anatomical images of anatomy. PET-CT investigations are of paramount importance in clinical oncology for diagnosing, staging and re-staging most cancers, monitoring response to therapy and planning radiotherapy treatment. Preclinical models imaged with PET-CT also form essential tools for the discovery of novel cancer therapeutics. To make quantitative assessments of PET-CT images contouring/segmentation of regions of interest are required. This can be contouring of cancerous lesions in clinical studies organs at risk during radiotherapy treatment planning and whole body organ wise contouring for biodistribution studies in the pre-clinical model. The ability to automatically segment PET-CT images is an open problem. Methods for image analysis developed in pre-clinical and clinical models are interchangeable; with knowledge transfer between the two domains common place. Manual approaches to identify and segment PET-CT volumes of interest are error prone.  The contouring of a whole body pre-clinical model can take several hours. The advantage of using a pre-clinical model to develop segmentation algorithms is the access to the ground truth anatomy. This project will build on an in-house developed automated segmentation tool for PET-CT imaging which utilizes deep learning (DL) artificial intelligence (AI). The student will assess the accuracy of the model in varying pre-clinical CT images which demonstrate variations in anatomy and radiopharmaceutical uptake; explore the use of human in the loop interactions to enhance accuracy and develop methods for knowledge transfer of the developed model to clinical datasets.

Aims and objectives

The aim of the project is to develop and evaluate deep learning models for automated segmentation of PET-CT images. Exploring the limitations wr.t training data variability and investigating solutions that increase model performance. These will include, reinforcement learning with HITL feedback and transfer learning for knowledge transfer between domains (i.e. pre-clinical to clinical imaging datasets). The aim of the project will be established via the following objectives:

Year 1 

1. To expand on our representative dataset of pre-clinical PET-CT images for training and validation of deep learning models. 
2. To enhance on our current deep learning model for automated segmentation of PET-CT images. The model is currently driven by CT images of normal anatomy. The student will incorporate state-of-the-art techniques and architectures to extend the model to utilize PET information and segment greater variations of normal/diseased anatomy with the inclusion of cancerous tumours. 
3. To evaluate the performance of the developed deep learning models on the collected dataset and identify limitations and areas for improvement.

Year 2 

1. To investigate and develop reinforcement learning techniques with human-in-the-loop (HITL) feedback to improve model performance.

Year 3 

1. To explore transfer learning techniques for knowledge transfer between pre-clinical and clinical imaging datasets, and evaluate the effectiveness of transfer learning on segmentation performance in clinical data. The student will work with clinicians based at Cardiff Universities PETIC and Velindre Cancer Centre (Dr Nick Morley). 
2. To provide insights and recommendations for improving the accuracy, robustness, and generalization capability of deep learning models for PET-CT image segmentation, with potential implications for clinical applications.

Materials and methods

In terms of physical sciences research areas, the project utilizes mathematical modelling and dynamical systems to develop and evaluate deep learning models. The project also uses quantitative omic data and image analytics to extract meaningful information from PET-CT images. The project recognizes the importance of data integration and trusted research environments, and the deep learning models developed in this project can be used to transform the way PET-CT images are analysed and interpreted. The project also adopts a human-centred AI approach, which focuses on developing models that are transparent, interpretable, and accountable to end-users. Overall, the project aligns well with the IPOCH research themes and areas, and has significant translational research implications for personalized medicine and cancer treatment.

Anticipated results

The project has translational research implications, as it aims to develop and implement deep learning models for PET-CT image segmentation and use transfer learning to examine how pre-clinical data analysis can inform imaging data in the clinical setting. The deep learning models developed in this project may also be integrated with decision support systems to assist clinicians in making more informed treatment decisions.

How to apply:

All applications should be submitted via the online application portal SIMS.

Further details on the application process can be found in the “how to apply” page with instructions for form completion here.

Online application portal is found at: https://www.cardiff.ac.uk/study/postgraduate/funding/phd-studentships-and-projects.

Along with the online application the candidate is asked to upload a covering letter, a CV, and two academic references (Reference Form Template). Transcripts of degrees and additional supportive documents can be provided at the interview stage.

Please complete the online application form by 5.00pm on Friday 30^th June 2023. If you are shortlisted for interview, you will be notified Friday 7^th July. Interview will be held during week commencing Monday 17^th July. Notification shortly after that certainly by the end of the month.

Cardiff University is committed to supporting and promoting equality and diversity and to creating an inclusive environment for all. We welcome applications from all members of the community, irrespective of age, disability, sex, gender identity, gender reassignment, marital or civil partnership status, pregnancy or maternity, race, religion or belief and sexual orientation. 

We welcome applications for full time study and from candidates with non-traditional academic backgrounds. For further information about, please contact us.

Engineering - Study - Cardiff University

Computer Science and Informatics - Study - Cardiff University

Medicine - Study - Cardiff University

Academic Criteria:

Candidates should hold or expect to gain a first-class degree or a good 2.1 (or their equivalent) in Engineering or a related subject. 

Essential skills: Highly numerate, excellent analysis and problem solving, effective written and oral communication, good project management and organization.

Desirable skills: Image processing, Matlab, python.