Oral Presentation 37th TROG Cancer Research Annual Scientific Meeting 2025

External validation of FET PET automated segmentation of glioblastoma using prospective trial data (#105)

Nathaniel Barry 1 2 , Jake Kendrick 1 2 , Pejman Rowshanfarzad 1 2 , Mubashar Hassan 1 , Roslyn J Francis 3 4 , Nicholas Bucknell 3 , Eng-Siew Koh 5 6 , Andrew Scott 7 8 , Martin A Ebert 1 2 3 , Robin Gutsche 9 , Norbert Galldiks 9 , Karl-Josef Langen 9 , Philippe Lohmann 9
  1. School of Physics. Mathematics and Computing, University of Western Australia, Perth, Western Australia, Australia
  2. Centre for Advanced Technologies in Cancer Research, Perth, Western Australia, Australia
  3. Department of Radiation Oncology, Sir Charles Gairdner Hospital, Nedlands, WA, Australia
  4. Medical School, University of Western Australia, Perth, Western Australia, Australia
  5. Radiation Oncology, Liverpool Hospital, Liverpool, New South Wales, Australia
  6. South West Sydney Clinical School, University of New South Wales, Sydney, New South Wales, Australia
  7. Tumour Targeting Program, Olivia Newton-John Cancer Research Institute, Heidelberg, Victoria, Australia
  8. Department of Molecular Imaging and Therapy, Austin Health, Heidelberg, Victoria, Australia
  9. Institute of Neuroscience and Medicine, Research Centre Juelich, Juelich, Germany

1. Background

FET PET is an emerging imaging modality for guiding radiotherapy for glioblastoma. Credentialing of the TROG 18.06 FET in glioma (FIG) trial demonstrated variability in clinician segmentation of the target volume on FET images. Automated segmentation offers an opportunity to harmonise segmentation for analysis of trial data and to facilitate the translation of study outcomes.

2. Aims

To validate an existing autosegmentation algorithm for defining the biological target volume (BTV) for glioblastoma on static FET PET images.

3. Methods

Fifty-two static FET PET images were available from post-surgical baseline and post-treatment follow-up for a prospective study of 24 glioblastoma patients treated with chemo-radiotherapy. The biological target volume (BTV) on each image set was segmented by a single nuclear medicine physician for comparison against BTVs identified via an independent deep learning network trained on a total of 699 FET PET scans and expert-defined ground truth BTVs. Volume correlation, Bland-Altman analysis, spatial overlap and distance metrics were used to assess agreement between the physician and automated segmentations. Correlations were able to be considered relative to repeatability via a small test-retest cohort.

4. Results

Correlation was excellent at baseline (R=0.97, p<0.001) and follow up (R=0.98, p<0.001) imaging. Median (interquartile range) of the Dice similarity coefficient (DSC) and surface DSC was 0.83 (0.68-0.92) and 0.95 (0.78-0.99), respectively. There was a slight bias towards the network underestimating the BTV relative to the physician. In two patients with lesions of the frontal lobe, the network segmented non-tumour uptake of the transverse sinus.

5. Conclusions

Although there is capacity to tune the network for greater generalisation, the network performance on our independent, external test cohort was comparable to that reported during internal validation. This suggests this network could underpin an analysis pipeline designed to automate the processing of FET PET data for guiding glioblastoma treatment.