Boundary Extraction for Imperfectly Segmented Nuclei in Breast Histopathology Images - A Convex Edge Grouping Approach

The detection of cell nuclei plays a significant role in automated breast cancer grading and classification. Although many algorithms for nuclei detection are present in contemporary literature, there is a general arduousness in automatically segmenting nuclei which have an inhomogenous interior and weak boundaries revealed by uneven staining. Such nuclei are common in high grade breast cancer cells. This paper presents an automated boundary extraction methodology for detecting the broken or missing boundaries of imperfectly-segmented nuclei in breast histopathology images. The images are first segmented using K-means clustering method, to retrieve the prospective nuclei regions which may contain these imperfectly segmented nuclei. Following this, a boundary extraction methodology based on the grouping of approximately convex boundaries is used to uncover missing edges and connect the gaps inbetween them. The study is focused on patchy and open vesicular nuclei which are common in high grade breast cancers and which normally pose a challenge for automatic segmentation techniques. Using a sample size of a 100 images of nuclei for this evaluation, the proposed method yielded sensitivity and specificity rates of 90% and 93% with average Hausdorff distance measuring 59. In comparison, the same three factors achieved by employing color-based K-means clustering technique amounted to 49%, 92% and 323, whereas color deconvolution yielded 85%, 69% and 373 and intensity-based segmentation returned 14%, 97% and 351.