Identification and measurement of blood vessel cross-sections in ultrasonic images
In atherosclerosis research non-invasive techniques, e.g. ultrasonic imaging is gaining importance. Analysis of ultrasonic images is complicated by a fairly high amount of noise, the presence of artefacts and incomplete edges. This project deals with determining presence or absence, position and size of blood vessel cross-sections in ultrasonic images. These blood vessels appear in the images as dark circles with a well-defined bottom, possibly a well-defined top and potentially very fuzzy "sides".
The mission is to develop and implement an algorithm that is able to determine if a blood vessel is present in the image, and if so determine its position in the image and its diameter. The job is somewhat complicated by the possible presence of "red herrings" in form of other projections i.e. images not containing circular cross-sections.
Robust general mosaic reconstruction of microscopic images
Putting a digital camera on a microscope opens a lot of possibilities for efficient imaging. However, there are situations when the resolution of a single image is not enough, or when a large specimen cannot be captured in a single image even at low optical magnification. This project deals with constructing a large image from a number of tiles, i.e. partial images. Commercial software for mosaic reconstruction typically require that partial images are taken in a known pattern, usually n x m images arranged in a grid. In the absence of a motorized microscope, this is typically not what the partial images look like. In our current setting the number of partial images may vary from two to about 15, they are not taken in a known order, they may not form a solid, filled area, they have different amount of overlap etc.
The image content may vary a lot, so the only assumption that can be made in this area is that they are focused, have a fair contrast and some amount of at least manually identifiable structural features. Usually the images contain at least some amount of object - background border, but the solution should not depend on this.
The mission is to develop and implement an algorithm that robustly is able to arrange a number of spatially unconstrained (well, they DO overlap) partial images into a whole image. Robustness, i.e. really finding a valid solution and identifying it as being valid (or clearly identifying failure to solve the job for a given input) is more important than speed. The reason for this is that with a robust implementation the resulting program can be made to run unsupervised.
About both projects
Preferred implementation is Java, Matlab is acceptable.