A Level-Set Approach to 3-D Segmentation of Lesions from T1-Weighted Spin-Echo DCE-MR Images

Student: Nikhil Rajguru

Active contours (or "snakes") are popularly used for segmentation of regions of interest in medical images. The algorithm developed through this research harnesses the properties of active contours in a level set approach to perform the segmentation of high-resolution heterogeneous in-situ tumors from the 3-D dynamic contrast enhanced magnetic resonance images (DCE-MRI). The regions of interest (ROIs) in this study were MDAmb231 tumor cells (the human breast carcinoma cell line) grown in the mammary fat pads of severe combined immuno-deficient (SCID) mice. This 3-D segmentation is a critical step towards achieving a data fusion between the functional images derived from the MRI acquisitions, and the reconstructed tumors in histology. Most segmentation techniques for DCE-MRI are directed towards brain segmentation, and were not well suited to the problem at hand. The edge sensitivity of the level-set function to the boundary of the lesion was enhanced by combining forces derived from the image including gradient vector flow. Since there were no standard markers (as in brain segmentation) that could be used, tissue characteristics could not be harnessed. Instead, the local probability near the lesion boundary was utilized for level-set evolution. Based on experimental validation, the proposed algorithm performed better than other applicable segmentation techniques based on active contours on the data at hand, and also the classical level-set approach.


Fig. 1: (a) Kass snake, (b) region growing + modified Kass, (c) GVF snake, (d) proposed algorithm.

This work was a collaborative effort with Prof. Robert J. Gillies (Dept. of Biochemistry and Arizona Cancer Center).