Nathaniel Barry
PhD candidate, UWA Medical Physics
nathaniel.barry@research.uwa.edu.au
Title: External validation of FET PET automatic segmentation on a single-centre, prospective dataset
Summary: The recent development of an automatic segmentation network trained on 699 18[F]-Fluoroethyl-L-tyrosine (FET) PET images of the brain from a German cohort has prompted an investigation into its performance on an Australian cohort. Here we present the results of an external validation of this network on 53 images from a previous, prospective study of 24 patients conducted at SCGH, which shows comparable performance. This novel, single-centre analysis constitutes a benchmark evaluation to be applied to a multicentre dataset from the ongoing FET PET in Glioblastoma (FIG) trial.
Branimir Rusanov
Radiation Oncology Medical Physics Registrar,
PhD candidate, UWA Medical Physics
branimir.rusanov@health.wa.gov.au
Title: Autoencoder Graph Neural Network for Anomaly Detection of Limbus Auto-contours
Summary: Limbus can rapidly segment a large number of organs, however, manual assessment of these structures can become overwhelming especially under time pressure. An automated AI model was developed to identify potentially anomalous/erroneous Limbus contours for all major organs. The model, still under development, has shown the ability to correctly associate out-of-distribution organ encodings with poor segmentation, thereby correctly identifying poor Limbus contours. The system, when complete, will run as a light-weight service to flag problematic contours and notify RO/RTs prior to beginning treatment planning.
Joel Noble
UWA Medical Physics Masters student
22903004@student.uwa.edu.au
Title: Generating Synthetic CT Using Deep Learning to Expedite Palliative Radiotherapy
Summary: This project aims to use AI to expedite palliative radiotherapy by eliminating the need for patients to attend simulation sessions. A deep learning model is being trained using pairs of diagnostic CTs (dCTs) that have been rigidly and deformably registered to their corresponding planning CTs (pCTs). If trained successfully, the model will be able to generate a synthetic CT (sCT) from a pre-existing dCT. The sCT will closely resemble the pCT that would otherwise have been acquired during a simulation session, thereby expediting the patient's treatment.
9 сен 2024
