Premium
SU‐E‐J‐139: Fuzzy Clustering Segmentation of Glioblastoma in T1‐MRI Imaging for Clinical Trials
Author(s) -
Cordova J S,
Schreibmann Eduard,
Hadjipanayis Constantinos G.,
Holder Chad A.,
Bansal Vivek,
Julio Sepulvedad,
Hasan Danish,
Guo Ying,
Fox Tim H.,
Crocker Ian R.,
Shu HuiKuo G.,
Shim Hyunsuk
Publication year - 2014
Publication title -
medical physics
Language(s) - English
Resource type - Journals
SCImago Journal Rank - 1.473
H-Index - 180
eISSN - 2473-4209
pISSN - 0094-2405
DOI - 10.1118/1.4888192
Subject(s) - contouring , segmentation , cluster analysis , concordance correlation coefficient , concordance , fuzzy logic , glioblastoma , computer science , artificial intelligence , nuclear medicine , pattern recognition (psychology) , volume (thermodynamics) , fuzzy clustering , mathematics , medicine , statistics , physics , computer graphics (images) , cancer research , quantum mechanics
Purpose: Generating brain tumor volume measurements in a reproducible and efficient manner is a difficult, yet necessary, component of response assessment. The purpose of this study was to adapt and validate a multilevel Fuzzy C‐means clustering algorithms for ROI tumor segmentation to allow consistent volumetric comparisons at multiple sites. Methods: Preoperative contrast‐enhanced T1W images from 37 glioblastoma cases were segmented using Fuzzy C‐means clustering‐based methods and compared to manually contoured volumes created by specialists. The same was done post‐operatively, using subtracted images to eliminate intrinsically T1‐hyperintense material (blood). Volume computations based on the MacDonald criteria were also used for comparison. Agreement and inter‐rater variability between volumes produced with each method was assessed by determining the concordance correlation coefficient (CCC). Results: The MacDonald criteria method had poor agreement (CCC=0.350–0.972) with manual contouring pre‐ and postoperatively, while the proposed semi‐automated methods exhibited very high agreement (CCC=0.839–0.995) with manual contouring before and after resection. Fuzzy C‐means clustering with three classes was the most robust semi‐automated method, showing better inter‐rater agreement than the MacDonald criteria method for both pre‐ (CCC of 0.990 and 0.975, respectively) and post‐operative cases (CCC of 0.983 and 0.576, respectively). Post‐operative inter‐rater agreement was significantly different between these methods (p < 0.001). Conclusion: The proposed semi‐automated segmentation methods allow tumor volume measurements of MR images in a reliable and reproducible fashion necessary for measuring treatment response in glioblastoma patients in multicenter clinical trials.