Determining the prognosis in an early stage of human cancer can be essential for the choice of optimal therapy. Digital image analysis of cell nuclei is a very useful tool to obtain quantitativeinformation for robust and reliable prognosis. A substantial number ofpapers have been published on the use of various texture analysismethods for diagnostic and prognostic work on human cancer, andmost of the studies are based on texture analysis of the gray levels in the images.
We will take another approach, and use a refined adaptivesegmentation method developed in this thesis to describe thestructures inside the cell nuclei images. The refined thresholdingmethod is spatially adaptive within each image, while its parametersare adapted to the histogram of each image. In a novel approach, weevaluate the characteristics of the segmented structures statisticallyto decide the prognosis per image, and finally a rule is formed toclassify each patient.
The data set analyzed consists of 134 patients with early ovariancancer. The problems with such small data sets is addressed, and a solution based on statistical bootstrapping isproposed. This gives a more robust estimate of the correctclassification rate (CCR) than the traditional single CCR estimatewould, and in addition gives a CCR uncertainty estimate.
Dividing the data set into two groups based on DNA-ploidy - effectively introducing a simple two-step classification scheme - substantiallyimproved the performance of the classification. Combining thestructural features extracted from the objects inside eachcell nucleus with the best statistical gray level feature - anadaptive entropy matrix feature from a previous study on the same material - further improved the correct classification rate, leading to a CCR close to 90\%.
In conclusion, the significant improvement in correct classificationrate obtained by combining the best statistical and structural texture features seems to hold a promise of very high CCRs, which would be immensely valuable in prognostic work on human cancers. This may betrue beyond the present data set, and possibly quite generally. But obviously some caution is called for, and more tests on differentand larger data sets should be performed.