Meeting Banner
Abstract #1521

Quantification of Necrosis in Animal Tumor Model Using K-Means Clustering of ADC

Louisa Bokacheva1, Khushali Kotedia, 12, Megan Reese1, Carl Le1, Jason Koutcher1, 3, Sean Carlin1, 3

1Department of Medical Physics, Memorial Sloan-Kettering Cancer Center, New York, United States; 2Baylor College of Medicine, Houston, TX, United States; 3Department of Radiology, Memorial Sloan-Kettering Cancer Center, New York, United States

Six athymic nu/nu rats bearing HT29 human colorectal xenograft tumors (700 mm3) were imaged at 7 T twice and diffusion-weighted images were acquired at five b-values between 0 and 900 s/mm2. After imaging, rats were sacrificed, tumors were excised and histologically analyzed to determine the necrotic fraction. ADC voxel maps were calculated using monoexponential equation. K-means clustering was applied to the ADC data pooled from all tumors and ADC maps were segmented into two clusters: (1) viable and (2) necrotic. The threshold ADC value between clusters was found to be 0.88x10-3 mm2/s. The fractional area of the necrotic cluster 2 correlated well with the histological necrotic fraction (R2 = 0.91), although for half of the tumors, necrotic fraction was slightly overestimated. The spatial agreement between the cluster maps and histological necrotic areas was better for tumors with larger contiguous necrotic areas than for tumors with multiple small necrotic regions.

Abstract PDF

Keywords

accuracy affect agreement although analyzed animal applied applying automatic background berry best better blue bottom built cancer central centroids close cluster clustering clusters coil college considerable considerably considered contiguous control correlated correlates correlation correspondence cutoff depend determined diameter diffusion distances distributions done doses drug empirically every excised explored expression female findings fitting fraction fractions function grown henning hind histogram histological histology home human identical identified identity implemented islands larger limb linear localized location maps masked matching matrix measured medical medicine memorial middle mixed model nearly necrosis necrotic often options oral outlined overestimated period physics pooled post previous produced quality quantification radiology rats receiving regression reports require requires response rigorous sacrificed sections segmentation segmented sensitive slice slices slightly smaller solution spatial spin stain stained studies subcutaneously table threshold tissue treatment tumor tumors turbo twice underestimation validation variety vary vehicle viable volume whole worse yielded