An External, Independent Validation of an <i>O</i>-(2-[¹⁸F]Fluoroethyl)-l-Tyrosine PET Automatic Segmentation Network on a Single-Center, Prospective Dataset of Patients with Glioblastoma.

Abstract

The goal of this study was to conduct an external, independent validation of an <i>O</i>-(2-[¹⁸F]fluoroethyl)-l-tyrosine ([¹⁸F]FET) PET automatic segmentation network on a cohort of patients with glioblastoma. <b>Methods:</b> Twenty-four patients with glioblastoma were included in this study who underwent a total of 52 [¹⁸F]FET PET scans (preradiotherapy, <i>n</i> = 23; preradiotherapy retest, <i>n</i> = 9; follow-up, <i>n</i> = 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_mean], lesion SUV_mean, and background SUV_mean) 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. <b>Results:</b> BTV and TBR_mean Pearson 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 cm³, LoA, -13.1-18.5 cm³; relative bias, 27.9%, LoA, -95.3%-151.2%) and deviations for TBR_mean were 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. <b>Conclusion:</b> 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.