Clustering-Based Image Registration of Tubule Candidates in Breast Histopathology Images

Tubule formation is a rudimental and vital parameter to be considered for cancer aggressiveness detections based on histopathology image data, in accordance with the Nottingham Histopathology Grading (NHG) system. In order to identify and measure tubule formation, here, the present study investigates the potential of clustering-based image registrations for tubule candidates’ identification. The elementary structures that constituted the signatures of the tubules: nuclei and central lumens are identified using K-means clustering. The breast histopathology image patches are harvested from the hematoxylin and eosin (H&E) stained breast Whole-Slide Images (WSIs). The stain unmixing method is utilized to normalize the image colors, and to compute the hematoxylin-stained channel (H-channel) and eosin-stained channel (E-channel). K-means clustering is applied on the H-channel to segregate the nuclei regions from the stain artifact regions or background. The same method is applied on the E-channel to identify the hemorrhage regions for segregation from the initial image. Also, the same algorithm is applied to the stainnormalized histopathology patch for the identification of lumen candidates’ regions. A simple heuristic axiom is designed and mechanized, prior to the pronouncement of precise lumen candidates’ cluster affirmations apart from the rest of the clusters. Through overlaying (summing) the nuclei and lumen candidates’ clusters on the same image frame, the candidates of tubules are registered from the reference input histopathology image patch.