Texture Analysis and Tissue Segmentation of Cryosection Images



Jennifer Casper and Tammy Williams
REU Summer 1998





INTRODUCTION
We investigated two methods to extract tissues in cryosection images based on texture patterns. We used the cryosection data from the National Library of Medicine specified for the Visual Human Project. The Visible Human Project was undertaken in order to create an anatomical atlas of the human body. The work presented here was done as part of the Research Experience for Undergraduates at the University of Wisconsin--La Crosse.


BACKGROUND ON THE DATA
We used the cryosection dataset of the male cadaver. The frozen male cadaver was axially sliced and imaged at one millimeter intervals throughout the length of the body. The images contain three bytes of information per pixel. This information includes the red, green, and blue values for each pixel. The cropped images that were processed for our work are 856 pixels by 495 pixels.


ABSTRACT OF THE RESEARCH PROJECT
Our research utilized texture as a means of isolating individual tissues from the cryosection images of a male cadaver obtained by the National Library of Medicine. Visual stochastic texture varies among different tissue types; thus, texture was chosen as the defining attribute for tissue segmentation. Texture, for our purpose, is defined as regular and irregular placement of color in an image. This work is a part of the Visible Human Project which will enhance the undergraduate study of anatomy. Students will be able to access and manipulate the cadaver data from the type of workstation found on the average college campus as opposed to access being limited to students at institutions which possess the high powered workstations for which most manipulation of these large data sets has previously been done. Simple, rapid segmentation techniques based on texture will aid in viewing individual organs and tissues on such hardware.

Two techniques for extracting tissues from the male cadaver dataset were compared. The first involved high-level decision-making algorithms. A cryosection image was traversed using an image template. Compared pixels were selected based on color value, position, and color variance. The second technique utilized an artificial intelligence approach. A feed-forward neural network provided a robust structure to detect texture. This customized neural net retained information about the color values and spatial location of a region in a cryosection image. A back-propagation algorithm trained the net by adjusting for optimal weights.

The decision-making algorithms produced satisfactory results for the extraction of multiple tissue types: fat, bone, and muscle. Most of the desired tissue was acquired although unwanted tissue was taken as well. The neural net produced better results in differentiating between a desired tissue type and all other tissue types in an image. The neural net generated images with sharper edges and fewer regions of unwanted tissue. The selection of unwanted tissue resulted from color similarities among tissues in the body and the absence of texture patterns in the images. The neural network offered a flexibility in distinguishing different tissues that was not apparent in the results of the decision-making algorithm. Although both methods use similar traversal techniques through a cryosection image, the neural network required more computation per pixel than the decision-making algorithms; therefore, the decision-making methods were more efficient. The neural network technique required more background processing and memory in order to train weights, but the two methods used comparable amounts of memory during the tissue extraction computations. The techniques were not used together, but improved results may be generated if the methods are combined or used in succession.

This work has shown that texture differences can be used successfully in isolating tissues from cryosection images. Two very different approaches to tissue segmentation were compared using minimal computing power on readily accessible workstations.

In conclusion, we have shown that texture is a very strong key for identifying tissue type, that it is possible to isolate tissue from cryosection images based on texture differences, and that this can be done without the use of parallel processors or the commitment of large amounts of memory.


TECHNICAL REPORT
Click here to view the postscript version of the technical report: "Texture Analysis and Tissue Segmentation of Cryosection Images"

RESULTS
The following links direct you to images resulting from the algorithms presented in the technical report.
Decision 4.0
Decision 5.0
Neural Net--Image 1245
Neural Net--Image 1350
Neural Net--Image 1737
Neural Net--Image 2095

AUTHORS
Jennifer Casper is completing her undergraduate majors in computer science and art at the University of Wisconsin--La Crosse. Jennifer participated in a two-year NSF REU robotics program at Colorado School of Mines during the summers of 1996 and 1997. She is currently a software developer at The Trane Company. Jennifer is planning to continue her education at the graduate level in the fields of robotics and artificial intelligence.

Tammy Williams is majoring in computer science and mathematics at Gustavus Adolphus College. Tammy participated in the REU program for the Visible Human Project at UW-L during the summers of 1997 and 1998. She is hoping to attend graduate school after her graduation in May of 1999. Her interests include computational theory and several branches of mathematics. Tammy enjoys spending her free time on the golf course and in competition as a member of the Gustavus golf team.

Questions/Comments about the research project:

Please send e-mail to Jennifer Casper or Tammy Williams.





LINKS TO MORE INFORMATION ON THE VHP

The Visible Human Project
Marching Through the Visible Man
Volume Rendering of the Visible Human Data


LINKS TO MORE INFORMATION ON TEXTURE

Representing Texture
Statistical Geometrical Texture Description


LINKS TO MORE INFORMATION ON NEURAL NETWORKS

Neural Networks
An Introduction to Neural Networks
What is a Neural Network?