The project deals with ultrasound images processing. It focuses on ultrasound images of premature infants’ brains.
Abstract
The project deals with ultrasound images processing. It focuses on ultrasound images of premature infants’ brains.
The Medical Problem
Hemorrhage in premature infants’ brains is quite common phenomena. The hemorrhage causes wide range of damages to the brain tissue, and may results in brain damage. Early diagnosis is crucial. Severe damage is not so hard to spot with naked eye (looks like a black hole in the image), while less severe damage would cause a slight change in the brain tissue, and would appear as a minor change in texture, in the ultrasound image. These areas are hard to spot with no other means but the doctor eye.
Our purpose was to write an algorithm that would help to locate the area with the less severe damage, in other words to give some indication that certain area in the brain is suspected to be injured.
The Algorithm
The algorithm should provide the ability to make a differentiation between different texture in the image. The assumption is that each texture contains different spatial frequencies. A simple Fourier transform would not do any good in this case, because it provides a global description of the signal energy as function of frequency. We are interesting in more detailed description, one that provides a local information on the signal energy as function of frequency.
To meet those requirements we used the 2-D Gabor Transform. The 2-D Gabor filters are appropriate for textural segmentation in several senses: they have tunable orientation and radial frequency bandwidth, tunable center frequency, and optimally achieve joint resolution in space and spatial frequency.
The local spatial frequency of an image is defined as followed: The image is multiplied by a function, that is somewhat centered around specified coordinate (X0,Y0) (as mentioned above, we used the 2-D Gabor function). The Fourier transform of the product is calculated. The result is the spatial frequencies around the coordinate: (X0,Y0). The process is repeated for each pixel in the image, which result in full spatial frequency description for the whole image. This is actually a convolution product between the Gabor function and the image.
The 2-D Gabor function is in a matter a fact a whole family of Gouse functions that differ one from the other by several parameters that controls the scale and the orientation of the Gouse. For each image we used S*K different filters, (S= number of scales; K= number of orientations) and got S*K different filtered images. It assumed that the local texture regions of the image are spatially homogeneous. Each pixel in the filtered image is characterized by two calculated parameters: the mean and the standard deviation for the immediate neighborhood (5 X 5 pixels). The same calculation is done for each of the filtered images, so for each pixel in the original image there is a 2*S*K parameters vector.
In this way each different reference texture was characterized with such a vector. We shell refer to them as the “reference vectors”. For each pixel in the examined picture, such a vector is calculated. The Euclid’s distance between it and the reference vectors is calculated also. The goal is to check to which of the reference vectors the current vector is closer (minimal distance). We assume that such closeness indicates on a similar texture in the original image. In this way we can get a differentiation between the different texture in the original image.
The algorithm was activated on a regular photo (the “brodatz” photo, which is a famous photo for texture analysis purposes), And got fine results (see below). In this picture there is a need to separate between 5 different texture.
The task was much more difficult with the ultrasound image, which is much more noised image, which contains a lot of “artifacts”. In this picture we tried to differentiate between three different texture:
(1) A normal brain tissue
(2) A damaged tissue (small holes)
(3) A suspected to be damage tissue (for that purpose we used hole surrounding)
The results is presented below:
The original Brodatz image The different textures as they are recognized by the algorithm
The original ultrasound image Texture recognition as to medicians
The different textures as they are recognized by the algorithm
NOTE: The different tissues in the original picture are marked according to Dr. Apelman’s instructions. She used more recent images, of the same brain to specify those areas.
Acknowledgments
We would like to thank our supervisor Chen Sagiv for her support and guidance throughout this project.
Also we would like to thank the Ollendorff Minerva Center Fund which supported this project.