Automated quantitative cellular analysis of immunohistochemically stained tissue sections

• Main Author: Demirkaya, Omer
• Format: Dissertation
• Language: English
• Subjects: engineering, biomedical; engineering, electronics and electrical
• ISBN: 9780591298055; 0591298058

An image analysis system for histochemically and immunohistochemically stained cells in histological tissue sections was developed. Image processing techniques were produced to preprocess, segment, label, and separate overlapping and/or touching cells.
Preprocessing and segmentation. Digital images of tissue sections were acquired using a line-scan CCD camera. Nonuniform illumination, DC bias, and spatially nonuniform gain of the CCD were corrected. Segmentation of macrophages and T-cells was performed in two stages: Kittler's, Otsu's and Kurita's thresholding algorithms were applied to roughly segment cells from background. Edge pixels of the region whose variance was larger were eroded if their value was larger than the mean by a specified fraction of the standard deviation. Segmentation of cell nuclei in hematoxylin-stained breast tissue was performed using local thresholding. Segmentation was evaluated by comparing computed and manually outlined cell areas. Macrophage areas computed at the first stage by Kittler's and Kurita's method differed significantly from those traced manually; those done by Otsu's method did not. Kittler's algorithm proved unsuitable for the segmentation of these cells, but the other two performed well when combined with the edge pixel erosion. For T-cells and cell nuclei, although automated and manual results agreed, large interobserver variability in manually outlined cell areas prevented the selection of a most accurate method. Inaccuracy and interobserver variability in manual tracing of cell clumps were also observed with phantom images.
Cell separation. The boundary curve of cell clumps was partitioned into homogeneous boundary segments using the minimum negative curvature points computed using scale-space analysis. This yielded a stable, noise-invariant partitioning process. Segments were then organized into clusters corresponding to individual cells using semantic knowledge (size and shape) and relational information (good continuity, compactness, proximity). The cells of eight clumps were manually counted by two trained observers. The observers' average did not differ significantly from the automated results. Comparison of computed and manual results showed a significant correlation, and regression analysis resulted in the unity curve. We conclude our method successfully separates cell clumps.