A multivariate method for ultrasound tissue segmentation for biomarker analysis of tumor growth

Date of Completion

January 2002


Engineering, Biomedical




The traditional means of analysis of tumor growth and morphology is to excise the tumor and study thin, histopathology slices under the microscope. This method requires a different subject for each time Estpoint and runs the real risk of missing some aspect of the tumor not collected in the slice. Ultrasound provide a means to study the tumor in vivo but the images can be hard to interpret. In this research, pixel intensity and contrast and entropy measurements of texture, derived from a cooccurrence function of the image, are used in a robust multivariate image segmentation algorithm to classify the tumor into viable and necrotic cells, reducing misclassification of tissue in the absence of reliable a priori information while allowing for a variable cost function for the type of misclassification. A nude mouse with an Hras tumor was used to establish the model and four nude mice with B16-F10 tumors were used to study the tumor growth over 14 days post cell injection. Histopathology images, one for the Hras tumor and one from mouse 3 of the B16-F10 tumors, were used to validate the segmented image. The multivariate method identified 73% of the necrosis using the mean pixel intensity and texture information while the intensity-alone method identified only 39% of the necrotic-associated pixels. ^