2006 Scientific Session Abstracts

Image Analysis for Neuroblastoma Classification: Hysteresis Thresholding for Nuclei Segmentation

http://www.bmi.osu.edu

Metin Gurcan, PhD1 (gurcan.1@osu.edu); Tony Pan, MSc1; Hiro Shimada, MD, PhD2, Joel Saltz, MD, PhD1, 1Department of Biomedical Informatics, The Ohio State University, Columbus, OH, 2Children’s Hospital Los Angeles, Los Angeles, CA

Context: Neuroblastoma is one of the most common cancers in infants and children. The current classification system, developed by Shimada et al., uses histopathology evaluation of tumor tissues. The Shimada classification partly depends on the type and state of cells, for instance, in relative counting of cells in mitosis or karryorrhexis. Therefore, identification and segmentation of cells and cell nuclei are important in computer-assisted analysis of neuroblastoma tumors.

Technology:  As part of an automated neuroblastoma image analysis system, we have developed a cell nuclei segmentation algorithm. The inputs to the algorithm are H&E-stained tumor samples digitized at 40x resolution using a commercial scanner (Aperio Scanscope, Vista, CA). The basic steps of the algorithm are color space decomposition, morphological reconstruction, hysteresis thresholding and post-processing. The output of the algorithm identifies the cell nuclei and delineates the boundaries of nuclei.
 
Design:  Hysteresis thresholding uses two thresholds: low and high. The threshold output consists of two sets of pixels: 1) all pixel values higher than the high threshold, 2) all pixels higher than the low threshold and connected to pixels higher than the high threshold. If only high threshold were used, only the highest intensity parts of the structures would be segmented. If only the lower threshold were used, then many false-regions would be segmented. Therefore, hysteresis thresholding provides a good compromise between these two extremes. For comparison, the segmentation algorithm was modified so that a constant threshold value was used instead of a hysteresis thresholding. The value of the constant threshold was selected as the average of the high (60) and low (45) hysteresis threshold values. The segmentation results with constant threshold value versus hysteresis thresholding were compared in terms of accuracy and its variation. The accuracy was tested by comparing the automated segmentation boundaries of the nuclei with those of the manual segmentation. For this purpose, 100 nuclei from different regions of neuroblastoma slides were manually segmented. The accuracy was determined as the average value of the overlap score between automated and manual scores. The variation was measured as the standard deviation of the overlap score for these 100 nuclei. Two different overlap scores were utilized: OS1 and OS2:

where M and A are manual and automated segmentation; , ,  denote the intersection, union, and the cardinal operations (i.e. the total number of “on” pixels in the segmented area), respectively.

Results:  The overlap scores with the constant thresholding were 85.76%±14.05% and 91.56%±10.39% for overlap scores 1 and 2, respectively. With the use of hysteresis thresholding, the overlap scores improved to 90.24%±5.14% and 94.79%±2.97%, respectively. 

Conclusion:  The use of hysteresis thresholding in combination with morphological pre-processing improved the accuracy for the segmentation of nuclei. Additionally, the variation in the segmentation outputs decreased.