Quantified is simplified; Treating the spatial entropy as continuous for prognostics of early ovarian cancer
Abstract
A substantial number of studies have proven that analysing the texture of DNA-specific stained cancer cell nuclei can provide robust and reliable prognostic information. Such information is important to make a qualified selection of the appropriate treatments for the patients. A recent texture approach based on adaptive features extracted from the class specific dual entropy matrix (CSDEM) has shown promising results. The approach used relatively coarse quantification of the entropy values to reduce overfitting. This quantification can easily reduce the performance of the approach, and will certainly require detailed domain knowledge in order to fully utilise its potential. We will in this study describe a method that uses the class specific entropy values in their continuous nature. The method uses an adaptive continuous discrimination function, based on density estimation, that is able to estimate the discriminative value of the entropies on a continuous scale. We have evaluated our method using statistical bootstrapping on a dataset containing about 38 000 cell nucleus images collected from 134 patients with early ovarian cancer. We achieve results that are consistently better than the quantified approach based on CSDEM, and our results are more easily obtained as domain knowledge requirements are reduced. Considering our method as a generalisation to the continuous domain, this is a good result that reinforces the promise of using the class specific entropies for prognostics of early ovarian cancer.