Analiza wpływu segmentacji jąder komórkowych w rozmazach Papanicolaou na pomiar wymiaru fraktalnego

W artykule badano wrażliwość metody pryzm trójkątnych (TPM - Triangular Prism Method) do estymacji wymiaru fraktalnego dla różnych rozmiarów jądra komórkowego w cytologii ginekologicznej wybarwionej metodą Papanicolaou. Zmiana obszaru analizy pozwala na zmniejszenie wpływu algorytmu segmentacji na wynik. Wykorzystanie kanału zielonego gęstości optycznej i zmian wymiaru fraktalnego dla par skal: 1-2 oraz 2-3 pozwala na otrzymanie wyników dla systemu klasyfikacji z małą wrażliwością na segmentację.

In the paper the influence of segmentation algorithms on estimation for the fractal dimension is analyzed. The Papanicolaou smears are very complex images and their automatic analysis is very hard. Segmentation algorithms of cell nuclei should support blurred and noised edges between cytoplasm and cell nucleus. The estimation of the cell nuclei image parameters is necessary, but the edge related parameters are not sufficient. The classification of the cells (correct/atypical) needs surface related parameters. Fractal based estimators are important for classification. The Papanicolaou images are colourful but only the green channel is important [9]. The TPM (Triangular Prism Method) is applied to the square area (2N+1 edge size) [5]. Multiple box selection variants occur and the multiple TPM analysis is applied and the mean value is calculated. The fractal dimension is calculated for a pair of scales (1-2, 2-3, 3-4). The correct and atypical cell nuclei are known and the analysis is separated. The histograms of difference between the known and reduced cell area are shown (Figs. 6-11). The atypical cells are less sensitive due to a larger size of the analysis area in comparison to the correct ones. Two scales (1-2) and (2-3) are useful, especially for smaller reduction parameter (erosion up to 9 pixels of original cell nuclei). Both scales are used in the classification system [9]. The fractal dimension changes are less than +/- 1%.