An External, Independent Validation of an-(2-[F]Fluoroethyl)-l-Tyrosine PET Automatic Segmentation Network on a Single-Center, Prospective Dataset of Patients with Glioblastoma.

in Journal of nuclear medicine : official publication, Society of Nuclear Medicine by Nathaniel Barry, Jake Kendrick, Pejman Rowshanfarzad, Ghulam Mubashar Hassan, Roslyn J Francis, Nicholas Bucknell, Eng-Siew Koh, Andrew M Scott, Martin A Ebert, Robin Gutsche, Keith George Ciantar, Norbert Galldiks, Karl-Josef Langen, Philipp Lohmann

TLDR

  • A study validating an automated segmentation network for [F]FET PET scans in patients with glioblastoma found excellent agreement with manual segmentation by an expert doctor, but noted limitations in volume estimation and erroneous segmentations.
  • The study suggests that further training on a representative local dataset is necessary for multicenter implementation and highlights the importance of careful evaluation and training of automated segmentation networks.

Abstract

The goal of this study was to conduct an external, independent validation of an-(2-[F]fluoroethyl)-l-tyrosine ([F]FET) PET automatic segmentation network on a cohort of patients with glioblastoma.Twenty-four patients with glioblastoma were included in this study who underwent a total of 52 [F]FET PET scans (preradiotherapy,= 23; preradiotherapy retest,= 9; follow-up,= 20). Biologic tumor volume (BTV) delineation was performed by an expert nuclear medicine physician and an automatic segmentation network. Physician and automated quantitative metrics (BTV, mean tumor-to-background ratio [TBR], lesion SUV, and background SUV) were assessed with Pearson correlation and Bland-Altman analysis (bias, limits of agreement [LoA]). Automated and physician segmentation overlap was assessed with spatial and distance-based metrics.BTV and TBRPearson correlation was excellent for all time points (range, 0.92-0.98). In 2 patients with frontal lobe lesions, the network segmented the transverse sinus. Bland-Altman analysis showed network underestimation of physician-derived BTVs (absolute bias, 2.7 cm, LoA, -13.1-18.5 cm; relative bias, 27.9%, LoA, -95.3%-151.2%) and deviations for TBRwere small (absolute bias, 0.03, LoA, -0.25-0.30; relative bias, 0.83%, LoA -14.27%-15.93%). Median Dice similarity coefficient, surface Dice similarity coefficient, Hausdorff distance, 95th percentile Hausdorff distance, and mean absolute surface distance were 0.83, 0.95, 10.94 mm, 3.62 mm, and 0.88 mm, respectively.Automated quantitative analysis was highly correlated with physician assessment; however, volume underestimation and erroneous segmentations may impact radiotherapy treatment planning and response assessment. Further training on a representative local dataset would likely be required for multicenter implementation.

Overview

  • The study aims to validate an automatic segmentation network for [F]FET PET scans in patients with glioblastoma, comparing it to manual segmentation by an expert nuclear medicine physician.
  • Twenty-four patients with glioblastoma underwent 52 [F]FET PET scans, and biologic tumor volume (BTV) delineation was performed by both the physician and the segmentation network.
  • The study assesses the accuracy of the automated segmentation network and the correlation between automated and physician-derived quantitative metrics.

Comparative Analysis & Findings

  • Pearson correlation analysis showed excellent agreement between automated and physician-derived BTV and tumor-to-background ratio (TBR) metrics across all time points.
  • Bland-Altman analysis revealed underestimation of physician-derived BTVs by the automated segmentation network, with a mean absolute bias of 2.7 cm and a relative bias of 27.9%.
  • The automated segmentation network performed well in quantitative analysis, with correlations ranging from 0.92 to 0.98, but showed limitations in volume estimation and erroneous segmentations.

Implications and Future Directions

  • The study highlights the importance of careful evaluation and training of automated segmentation networks before their implementation in clinical practice, particularly for radiotherapy treatment planning and response assessment.
  • The study suggests that further training on a representative local dataset may be necessary for multicenter implementation of the automated segmentation network.
  • Future studies could investigate the use of advanced deep learning methods and computational resources to improve the accuracy and robustness of automated segmentation networks.
Read Full Article