Computing Reviews
Today's Issue Hot Topics Search Browse Recommended My Account Log In
Review Help
Search
Brain tumor segmentation from multimodal magnetic resonance images via sparse representation
Li Y., Jia F., Qin J. Artificial Intelligence in Medicine73 1-13,2016.Type:Article
Date Reviewed: Jul 13 2017

The segmentation of brain tumors in magnetic resonance imaging (MRI) is clearly not only an important image processing task, but one in which the achievement of high accuracy can be life-changing for many. Although MRIs are performed in a number of different modes chosen specifically to increase contrast between key tissue types, the task of labeling voxels as tumor tissue is difficult. The fusion of data from several modes has the potential to allow accurate voxel labeling, but suitable algorithms must be devised.

In this paper, Li et al. report an automated algorithm for brain tumor segmentation that uses information from T1 (short time window), T2 (longer time window), T1C (T1 with contrast enhancement), and flair (very long time window) MRI scans to label image voxels. Five labels are used: normal tissue, edema, non-enhancing gross abnormalities, enhancing tumor core areas, and necrosis. Classification is done using a sparse representation-based classifier (SRC), first introduced for the face recognition problem by Wright et al.

The voxels from each of the four modes are concatenated into a single vector. This vector is translated into a sparse coding by reference to the union of five “dictionaries,” each generated from sample data for one of the five classes of tissue to be assigned. The coding scheme is based on the documented SRC method, except that information is included to consider the adjacency of voxels. The principle here is that the various classes of tissues generally exist in connected regions. Identification of the minimum residual classification is done by a Markov random field energy minimization process, with the near-optimal solution given by a mincut process. Many steps in the process are stated only by reference to other works, which enhances readability but will require extensive use of references to fully understand the method. Figure 1 is a clear and helpful diagram of the method; figures like this would be welcome additions to almost any paper.

The authors demonstrate their results by comparing performance on images used in a recent challenge in this area: the NCI-MICCAI 2013 Challenge on Multimodal Brain Tumor Segmentation. They report accuracy comparable with the leading algorithms in that challenge, with an overall performance that would rank second. Representative results are shown in image form, as well as graphically.

The discussion section of the paper is noteworthy. It explores the importance of setting parameters on algorithm performance, especially patch size and sparsity factor. Once again, this is a practice that most papers could benefit from and that is especially helpful to practitioners. The paper is well written and well referenced, and effectively presents a high-performing segmentation method for multimodal MRI data.

Reviewer:  Creed Jones Review #: CR145420 (1711-0752)
Bookmark and Share
 
Segmentation (I.4.6 )
 
 
Health (J.3 ... )
 
 
Markov Processes (G.3 ... )
 
Would you recommend this review?
yes
no
Other reviews under "Segmentation": Date
Knowledge-based interpretation of outdoor natural color scenes
Ohta Y., Pitman Publishing, Inc., Marshfield, MA, 1985. Type: Book (9789780273086734)
Jun 1 1986
Image segmentation and uncertainty
Wilson R., Spann M., Research Studies Press Ltd., Taunton, UK, 1988. Type: Book (9780863800672)
Feb 1 1990
Surfaces in range image understanding
Besl P., Springer-Verlag New York, Inc., New York, NY, 1988. Type: Book (9789780387967738)
Dec 1 1989
more...

E-Mail This Printer-Friendly
Send Your Comments
Contact Us
Reproduction in whole or in part without permission is prohibited.   Copyright 1999-2024 ThinkLoud®
Terms of Use
| Privacy Policy