Segmentation method comparison for baseline [18F]FDG PET-CT in follicular lymphoma patients

Abstract

Background Follicular lymphoma (FL) is the second most common subtype of non-Hodgkin lymphoma. Currently, [18F]FDG PET-CT is used for staging, response evaluation, and remission assessment. While advances in quantitative PET-CT are promising for prognostic assessment, they depend on reproducible tumor delineation. Various segmentation methods have been proposed, but their application to FL PET is less established, despite known differences in uptake patterns across lymphoma subtypes. This study aims to evaluate the performance of several single-threshold and multi-threshold methods for FL [18F]FDG PET-CT lesion segmentation on segmentation quality, interobserver variability, and ease-of-use. Methods Baseline PET-CT data of 25 second-line FL patients from the HOVON110 trial and 12 first-line FL patients from the PETAL trial were selected. Two observers applied 13 different semi-automatic methods, of which six used a single threshold and seven combined thresholds (multi-threshold). Methods include, SUV threshold methods, an AI-based method, majority vote and lesion-based selection methods. The segmentation process comprises four steps: step 1 and 2 involved generating a preselection, while step 3 and 4 applied an automatic method followed by manual adjustments. To assess segmentation quality, both observers gave a score (1–3) ranging from undersegmentation to oversegmentation. For interobserver variability, the difference in total metabolic tumor volume between observers was determined. The ease-of-use was assessed based on manually added and removed volume in step 4. Results A total of 962 segmentations were made by two observers. Differences in results between the methods were limited across all characteristics, indicating an overall satisfactory performance of all methods. The multi-threshold method scored better for segmentation quality in comparison to single-threshold methods, indicating less under- or oversegmentation. The single-threshold method SUV4.0 demonstrated lower median (0.3 mL) and inter quartile range (2.0 mL) concerning interobserver variability in comparison to lesion-based methods. Conclusion Among the single threshold methods, SUV4.0 is preferred regarding easeof- use, observer variability and segmentation quality. While the multi-threshold lesionbased methods showed the a higher segmentation quality, SUV4.0 has the benefit of easy implementation, wide availability and is in-line with the currently set benchmark for lymphoma PET analysis. We identified SUV4.0 and a lesion-based method as the candidate methods preferred for further clinical performance evaluation.