The goal of this project is to develop stochastic image analysis and computer vision techniques to obtain accurate volume computation of tumor. Due to complex structures of different normal tissues such the gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF) and abnormal tissues such as cyst, edema, necrosis and scar in the MR brain images, robust measurement of tumor is challenging. We have developed multi-resolution fractal stochastic feature extraction, unsupervised clustering and information theoretic approach to obtain robust tumor segmentation. Currently we have applied our tumor and non-tumor segmentation technique on different datasets. In order to compare efficacy of our result with other state-of-art techniques we consider 10 low grade patients with 1492 MRI slices from publicly available MICCAI BRATS2012 dataset. We obtain promising Dice overlap results (on an average 70%-80% for these patients. For a second dataset from our clinical collaborator, Children's Hospital of Philadelphia, we obtain on an average 63% Dice overlap score.
Figure 1. (a) T1 image, (b) T2 image, (c) FLAIR image, (d) Segmented Tumor by our algorithm (e) Tumor mask (original tumor outlined by radiologist).
The ODU Automatic Tumor Detection Tool has been developed in order to provide a solution to the detection of brain tumors. This tool is made available for evaluation. In order to obtain the tool, please complete the form below and we will provide you with a download link. To learn more about the tool, please see the video demonstration below. Please be sure to read the documentation before using the tool.
Video demonstration of automatic tumor segmentation tool
This project is partially sponsored by and/or in collaboration with the following: NIH, CHOP