Program Information
Comparison of Parameter Calculation Algorithms for DCE-MRI: Results From a Multi-Institutional Study
R Ger1,2*, A Mohamed1,3 , M Awan4,5 , Y Ding1 , K Li6,7 , X Fave1,2 , A Beers8 , B Driscoll9 , H Elhalawani1 , D Hormuth10 , P van Houdt11 , R He12 , S Zhou1,2 , K Mathieu1 , H Li1,2 , C Coolens9,13,14 , C Chung1,9 , J Bankson1,2 , W Huang6 , J Wang1,2 , V Sandulache15 , S Lai1 , R Howell1,2 , R Stafford1,2 , T Yankeelov10 , U van der Heide11 , S Frank1 , D Barboriak16 , J Hazle1,2 , L Court1,2 , J Kalpathy-Cramer8 , C Fuller1,2 , (1) The University of Texas MD Anderson Cancer Center, Houston, TX, (2) The University of Texas MD Anderson Cancer Center UTHealth Graduate School of Biomedical Sciences, Houston, TX, (3) University of Alexandria, Alexandria, Egypt, (4) Case Western Reserve University, Cleveland, OH, (5) University Hospitals, Cleveland, OH, (6) Oregon Health & Science University, Portland, OR, (7) The International School of Beaverton, Beaverton, OR, (8) Massachusetts General Hospital, Charlestown, MA, (9) Techna Institute, Ontario, Canada, (10) The University of Texas, Austin, TX, (11) The Netherlands Cancer Institute, Amsterdam, the Netherlands, (12) United Imaging Healthcare America, Houston, TX, (13) University of Toronto, Ontario, Canada, (14) Princess Margaret Cancer Centre, Ontario, Canada, (15) Baylor College of Medicine, Houston, TX, (16) Duke University Medical Center, Durham, NC
Presentations
TU-AB-601-4 (Tuesday, August 1, 2017) 7:30 AM - 9:30 AM Room: 601
Purpose: To test the inter-algorithm variability of quantitative metrics Ktrans and ve, which have been correlated with patient outcomes in head and neck squamous cell carcinoma (HNSCC).
Methods: Eleven different algorithms which implement either a Tofts-Kermode or extended Tofts pharmacokinetic model were compared. Digital reference objects (DROs) with known Ktrans and ve values were used to evaluate algorithm performance at different noise levels using absolute error and stratified permutation tests. DCE-MRI images of 15 HNSCC patients obtained at three time points during chemoradiotherapy were used to identify trends in Ktrans and ve over time. To determine if trends agreed across algorithms, Ktrans and ve were extracted from six regions of interest and evaluated using linear mixed effects models, Krippendorff’s alpha, and Spearman correlations.
Results: Average error of the measured parameters was less than 3% in the noiseless DRO. In the noisy DROs, the average error increased, but the measured Ktrans and ve values were ordered correctly for 86% and 84% of the DRO-algorithm combinations, respectively, according to stratified permutation tests (p<0.05). Values across algorithms were significantly different according to Wilcoxon rank-sum tests. This indicates that absolute metrics may not be used, but relative metrics may be valid for a given algorithm. In the patient data, the algorithm was a significant factor in most linear mixed effects models even when grouping algorithms by pharmacokinetic model. All Krippendorff’s alpha values were below the threshold for agreement indicating that algorithms did not consistently classify patients. A majority of algorithms produced a statistically significant Spearman correlation in ve of the primary gross tumor volume with time; no other correlations were found across a majority of algorithms.
Conclusion: Our results demonstrate significant limitations in combining data and results from different algorithms. Careful algorithm quality assurance must be carried out before comparisons among algorithms are performed.
Funding Support, Disclosures, and Conflict of Interest: 1- NIH/NIDCR 1R01DE025248-01/R56DE025248-01. 2- NCI/Big Data to Knowledge (BD2K) Program (1R01CA214825-01). 3- NIH/NCI Head and Neck Specialized Programs of Research Excellence (SPORE) Developmental Research Program Career Development Award (P50CA097007-10). 4- Rosalie B Hite Fellowship.
Contact Email: