Ultrasound image processing is a vast, complicated field, which involves coping with a large variety of problems.
Abstract
Ultrasound image processing is a vast, complicated field, which involves coping with a large variety of problems. In this project, an attempt was made to quantify the homogeneity level of the texture presented in an ultrasound image. To achieve this, different tools, which reflect different approaches to the problem, were checked. Finally, three directions were chosen: a statistical measure, a fractal measure and a measure which tests the image’s centers of mass (geometric and energetic). The effectiveness of these directions will be determined by medical experimentation.
As the project reaches its end, the medical staff will receive a basic tool for numeric characterization of textures. In the future, after adding new measures to it and building a database of typical values, we hope the computerized system will be able to assist in diagnostics.
The problem
Today, analysis of ovarian tumors through ultrasound images is performed by doctors, according to what they see in the image. Therefore the analysis is subjective, cannot be quantified and is bound to be different from picture to picture.
This causes a problem in comparing the analyses made by different doctors in different cases, thus making the act of diagnosis according to past experience much harder.
The solution
Our project aims to supply the medical staff at Bnei Zion Hospital with numerical measures for texture analysis of tumors.
After screening several algorithms for texture analysis, we chose 3 measures. These measures will be evaluated by the medical staff, and hopefully will assist in tumor diagnostics.
The first measure is a statistical analysis of the image (or a part of it). The calculated values are the image’s mean value, standard deviation, skewness and kurtosis. The last two values characterize the images histogram, which is also produced. We believe this set of measures will serve well in analysis of small, relatively uniform, textures.
The second measure is a calculation of the image’s fractal dimension. Recent studies have shown that the fractal dimension of an image correlates with the human perception of texture. In order to calculate this measure, we used the SDBC (Shifting Differential Box Counting) technique.
An in depth description of the fractal dimension and the algorithm we used appears in the project presentation and project report.
The last measure supplied within this project is the analysis of the image’s centers of mass. The distance between the geometric and energetic centers of mass is calculated, and is monitored as we look at the image in increasingly smaller “shells”.
From the tests we ran it seems this algorithm can supply a numerical indication of the image’s asymmetry, which is known to play a part in distinguishing benign tumors from malignant ones.
Figure 1 – Block diagram of the project
Tools
This project was developed using Matlab 7.1 and images from the Bnei Zion Ultrasound Institute image bank.
The Final Product
After having determined the algorithms for texture analysis, we created a user interface which will enable the medical staff to evaluate (and later use) these algorithms. The GUI enables the user to choose a region of interest inside the image and perform different calculations on it. The results of these calculations can then be exported to an Excel sheet for future use. Also, a ruler is available, for measuring sizes in the image.
Figure 2 – Main screen of the user interface
Figure 3 – Center of mass analysis through user interface
Conclusions
The conclusions from this project are not final yet, since the results are yet to be evaluated by the medical staff. However, from tests we ran on a variety of images, we believe the 3 measures supplied in the project may truely assist doctors in the analysis of ultrasound images.
Acknowledgment
We are grateful to our project supervisors, Dr. Zvi Leibovitz and Mr. Eli Appelboim, for their help and guidance. Throughout the project we received all the possible patience and support.
In addition, the staff at the VISL lab was very welcoming and helpful.
We are also grateful to the Ollendorf Minerva Center Fund for supporting this project.