Providing a robust and reliable estimation of a patient's prognosis is necessary to make a qualified selection of the appropriate treatment for that patient. Digital image analysis of cancer cell nuclei is useful to make such estimation. In particular, texture analysis of the DNA organisation of nuclei has through a substantial number of studies proved to provide quantitative information of prognostic relevance.
Most previous studies have used the first, second or higher order statistics to estimate the prognosis, i.e. applied statistical texture analysis. We will in this study take a different approach where we attempt to exploit the internal structure of DNA-specific stained nuclei. In our novel approach, we apply a novel, refined adaptive segmentation method to extract small dark and bright structures within the nuclei, and estimate the spatial entropy of the dark or bright structures of each nucleus based on the area of the segmented objects. Finally, we will use the spatial entropies to obtain some very few, but powerful novel adaptive texture features by adaptively estimating the discrimination value of each spatial entropy using the combined knowledge of all relevant spatial entropies of all nuclei across a number patients.
We have analysed our novel approach on a dataset containing 134 patients with early ovarian cancer when using a proper evaluation method based on statistical bootstrapping. The results are very promising. Our method performs significantly better than the previously most promising method based on texture analysis. Moreover, it performs consistently at least about equally well as all other approaches based on image analysis. Combining the best feature of our novel approach with a single other feature, we also obtain the best performance among all approaches based on image analysis.
If selecting a subset of the dataset based on a set of predefined criteria unrelated to digital image analysis, our novel approach attains a correct classification rate of 84 \%. This facilitate to a two-step recognition system. Again, our novel approach is consistently better, perhaps also significantly better, than all other approaches based on image analysis.
In conclusion, our novel approach seems to hold a promise of reliable estimation of the prognosis, which is necessary to make a qualified selection of the appropriate adjuvant treatment. Due to a very low dimensionality and the use of proper performance estimation, we expect that our approach will generalise well on an independent validation dataset. Moreover, because of the combination of high adaptivity in all stages of our approach and an addressed concern for the overfitting problem, we expect relatively good generalisation beyond the case under study. Nevertheless, caution must be called for, and new proper tests must as always be performed in the case of generalisations.