FEIT Research Project Database

Use of Deep Learning computational methods to characterise cancer burden in PET imaging

Project Leader: Leigh Johnston
Collaborators: Nick Hardcastle (Peter MacCallum Cancer Foundation), James Korte (Peter MacCallum Cancer Foundation), Price Jackson (Peter MacCallum Cancer Foundation)
Primary Contact: Leigh Johnston (l.johnston@unimelb.edu.au)
Keywords: cancer; imaging; machine learning; medical image analysis
Disciplines: Biomedical Engineering

PET imaging with tracers that target prostate-specific membrane antigen (PSMA) or the somatostatin receptor (SSTR) are currently be used to monitor disease volume, avidity, and presence of multiple metastatic sites which are implicated prognosis and inform treatment choices in patients with prostate and neuroendocrine cancers, respectively. While it is common to report measures for individual lesions in the form of maximum standardised uptake value (SUVmax), whole body metrics of tumour burden require manual delineation which may not be practical to perform for all patients. This project will explore the use of deep learning methods automate whole body assessment of PET images obtained from the tracers used for these diseases, 18F-FDCPyL (PSR) and 68Ga-DOTA-octreotate (GaTate).
Research will extend from our previous work with Deep Learning image analysis using Convolutional Neural Networks (CNNs) to identify structural anatomy in CT images. In this project we will optimise a 3D image segmentation network based on the U-Net CNN to delineate tumour boundaries on PET/CT images. The network will be trained to discriminate between regions of high PET tracer uptake due to malignancy from areas of normal physiological retention. The work is suited towards those with a background in either computer science, biomedical or electrical engineering or medical physics.

Further information: Please contact Dr James Korte (James.Korte@petermac.org) for further information about this project.