An Automatic Segmentation Algorithm for Colonic Glandular Lesions

Migyung Cho, Hyekyung Lee, Hwan Gue Cho

http://doi.org/10.5626/JOK.2018.45.6.554

Adenoma and adenocarcinoma of the colon are one of the most common tumors, and diagnoses are based mainly on the structural appearances and changes in cell morphology of the glandular structures. Each diagnosis is based on subjectivity and objectivity of each pathologist, and many studies are under way to extract meaningful features from the glandular structure for better objective results and reproducibility. Passive segmentation of glandular cells to extract structural features is a labor-intensive task performed over many hours and with some difficulties. These problems require an automated approach to quantify the shapes of glandular cells. In this paper, we have developed an algorithm for segmentation of glandular cells to quantify their shapes in the benign and initial stages of deformation signifying the onset of disease. The algorithm sequentially applies adaptive thresholds obtained by k-means clustering and obtains binary images by thresholding and filtering methods. We extract boundary information from binary images and combine several boundary information, and then we search for glandular cells, both in the outward direction and inward direction from the boundary information. Applying the proposed algorithm to clinical images showed more than 95% accuracy. In addition, it is a very practical algorithm because it is much faster than the level-set based algorithms.