|dc.description.abstract||St. Olavs hospital has supplied a dataset of 2703 tissue samples from the tumor
periphery from approximately 900 patients organized on tissue microarrays
(TMA). In this project we wish to examine all these tissue samples with image
processing to determine if second harmonic generation microscope images of
tissue can improve classification of cancer type (I, II, III) or in other
words, cancer aggresiveness. This thesis documents methods which automates the
microscope imaging of TMA and show how images can be correlated to clinical
data. Datamining methods can then be used on this dataset to look for patterns
which can be used in classifcation.
Automated microscope scanning is easy in consept, but the implementation
depends on many aspects of the experimental setup. Some of the aspects
discussed in this thesis are:
- Develop image analysis algorithms that are robust to experimental variations.
- Handle systematic errors like intensity variation and rotation between
scanning raster pattern and stage coordinate system.
- Automatic stitching of regular spaced images with little signal entropy in
- Adjusting z-plane tilt for large area samples with micrometer precision.
- Interfacing with commercial Leica software.
The focus of this thesis is on TMA and the experimental setup with a Leica
SP8 microscope, but some the aspects listed above are not unique to this context
The conclusions are:
- Large area scans should adjust specimen plane to be at even distance to the
objective to be time effective and avoid out of focus images.
- Using heuristics/constraints improves the reliability to automatic stitching
algorithms, failing gracefully on images with little entropy in overlap.
- Leica LAS version X 126.96.36.19920 have limited support for automatic
microscopy, but it's possible to work around limitations to leverage fully