Abstract
503
Objectives: 18F-DCFPyL, a small molecule inhibitor of the prostate-specific membrane antigen (PSMA), has shown significant promise in the detection of primary and metastatic prostate cancer (PCa) lesions. The objective of our study is to assess the reproducibility, reliability and relationships between quantitative textural imaging features in 18F-DCFPyL PET images.
Methods: 25 patients with National Comprehensive Cancer Network high- or very-high-risk primary PCa planned for surgery were imaged with 18F -DCFPyL at 60 min post injection. The prostate tumors were delineated using 6 different segmentation methods: manual delineation and gradient-based semi-automatic segmentation both performed by an experienced PET reader (MD), and 4 threshold-based segmentations ranging from 30% to 60% of SUVmax. Images were resampled to isotropic cubic voxels for consistency. Two quantization methods (uniform and Max-Lloyd) with 7 different gray-levels (4 to 256) were considered for 2nd-order textural features extraction. In addition to 5 conventional quantitative uptake measurements (SUVmax/peak/mean, tumor volume (TV), total lesion activity (TLA)), a total of 87 radiomics features were extracted for each patient/segm./quant. including: 11 first-order, 9 morphological, 27 Gray Level (GL) Co-occurrence Matrix (GLCM), 12 GL Run Length (GLRLM), 13 GL Size-Zone (GLSZM), 5 Neighborhood Gray Tone Difference (GLTDM) and 10 moment-invariant (MI) features. Spearman rank-correlation analysis was performed for 7 quantization levels in each of the two quantization methods to determine reliable and practical gray levels. To inspect the reproducibility of features across 6 segmentations and 7 gray levels of 2 quantization methods, we quantified the intra-class correlation (ICC). The statistical relationship between all 92 features were also explored based on the optimum gray level and manual segmentation.
Results: SUVmax as derived from primary tumors of 25 patients ranged from 5.6 to 51.8 (15.8±11.5). TV ranged from 0.5 to 61cc (7.3cc±11.8cc). Uniform quantization showed consistent Spearman correlations for gray levels >=64, while significant details on uptake heterogeneity were lost for levels <=32. Max-Lloyd quantization had a similar pattern of loss for <=32 levels. Feature rank correlations against SUVmax and TV exceeded 0.7 for 37% and 0.9 for 13% of the feature pairs, indicating limited complementary information for the latter features. Many features were highly correlated with one or more of TV, TLA, GT2 (TLA of uptakes>2), homogeneity and entropy (GLCM) (p-values<1e-6), which we subsequently focused on. Correlations of radiomic features against SUVmax and TV tend to decrease substantially with ranges of increasing SUVmax and volume. This is because the partial volume effect dominates quantification of smaller tumors, while larger ones tend to exhibit higher anatomical and physiological complexity in the uptake spatial distribution. 13 features (6 SUV-based, 3 morphological and 3 GLSZM) had ICC > 0.6 for all 6 segmentations. Reproducibility increases substantially considering only manual and gradient-based segmentations, yielding 29 features with ICC > 0.8. Radiomic features exhibiting both high quantization ICC (> 0.85) and high segmentation ICC (> 0.85) include RLN (GLRLM), entropy (GLCM), ZSN and ZSV (GLSZM), J3 (MI) and GT2.
Conclusion: Quantization by 64 gray-levels is sufficient to capture heterogeneity information. Threshold contouring significantly impacts reproducibility of features, and user-guided or more advanced segmentations are recommended for more reliable radiomics analysis. Entropy (GLCM), ZSV and ZSN (GLSZM) exhibit high reproducibility and reliability in this study of 18F-DCFPyL prostate PET as well as past studies on 18F-FDG, indicating their high reproducibility across different tracers, and are recommended for investigation of their possible prognostic or predictive value. Research Support:
In this issue
Jump to section
Related Articles
Cited By...
- No citing articles found.