Program Information
Modeling of the Dice Coefficient for PET Segmentation Studies
R McGurk1*, V Smith2, J Bowsher3, J Lee4, S Das3, (1) Medical Physics Graduate Program, Duke University, Durham, NC, (2) Department of Biostatistics, University of North Carolina at Chapel Hill, Chapel Hill, NC, (3) Department of Radiation Oncology, Duke University Medical Center, Durham, NC, (4) Université Catholique de Louvain, Brussels
TH-C-WAB-8 Thursday 10:30AM - 12:30PM Room: Wabash BallroomPurpose:
The Dice coefficient (DC) is a common accuracy metric in PET segmentation studies that is a measure of overlap between the segmented volume and ground truth. This work presents a methodology to model DC as a function of object shapes, sizes, contrasts, noise levels and filters.
Methods:
Five spherical volumes (1.1-26.5ml) and two irregular volumes (16&32cc) were imaged for 1, 2 and 5 minutes at high (~8:1) and low (~4:1) contrast. A Gaussian filter (5mm FWHM (G5)), and a bilateral filter (7mm spatial FWHM with adaptive intensity range kernel (B7)) were applied with and without a 3 mm FWHM Gaussian pre-smoothing step (filters G3G5 and G3B7). Adaptive and thresholding at 40%-max was used for segmentation and DC values produced by comparing the segmentation with a ground truth defined from a high-resolution CT. Generalized estimating equations (GEE) were used to fit models describing DC as functions of object size, contrast, scan duration, filter and segmentation method and compared with parameters estimated from ordinary least-squares fitting.
Results:
DC was most affected by sphere size (13mm vs.37mm,∆DC=-0.22,p<0.0001), followed by contrast (4:1 vs.8:1,(∆DC=-0.07,p<0.0001), then scan duration (1 vs.5min,∆DC=-0.06,p<0.0001). DC values were most improved with a combination of G3G5 filter and ADP segmentation (G3G5&ADP vs G5&40%,∆DC=+0.04,p<0.0001) for spheres. Scan duration had the biggest effect for irregular shapes (1 vs.5 min ΔDC=-0.07,p<0.0001). No combination of filter and segmentation method produced significantly higher DC values than those for the G5 filter and 40% thresholding. The significance of explanatory variables changed in five cases for spheres, and three for irregular shapes between the GEE and OLS model fits.
Conclusion:
DC was most affected by object size for spheres and scan duration for irregular volumes. The GEE framework accounts for the bounded, correlated and heteroscedastic nature of DC values, and are recommended for the analysis of DC.
Funding Support, Disclosures, and Conflict of Interest: Ross McGurk is supported by a New Zealand Bright Futures Top Achiever Doctoral Scholarship, Grant Number DKUX09001.
Contact Email: