Colon Biopsy Classification Using Crypt Architecture

Assaf Cohen, Ehud Rivlin, Ilan Shimshoni, Edmond Sabo





In this paper we introduce a novel method for the detection of diseases in biopsies that contain glandular structures. Most approaches proposed in the literature try to classify the biopsy without analyzing its basic elements (such as the nuclei and the glands). The proposed method differs in that it is based on the architecture of the glands in the biopsy and the analysis of each pixel. We demonstrate our novel, three-step method on the task of classifying colon biopsies. First, as described in our previous work, we create a pixel-level classification image, segment the crypts (the glandular structures in the biopsy) using it, and remove false-positive segments. Next, we calculate the crypt architecture using Delaunay triangulation on the crypt centroids and use this architecture to retrieve those crypts that were incorrectly removed in the first step. In the final step, we use the segmented crypts to construct a more accurate architecture and classify each triangle as healthy or cancerous using the classification of the crypts as healthy or cancerous. The method was tested on 54 colon biopsy images: 109 healthy sub-images containing 4944 healthy crypts and 91 cancerous sub-images containing 2236 cancerous crypts. It achieved 92% accuracy in crypt classification and 94% in biopsy region classification.



We present all the results and ground truth (crypts and region segmentation to benign and cancerous) of all the images on which the algorithm was tested. Download. You are free to use this dataset but please cite:


A Cohen, E. Rivlin, I. Shimshoni, and E. Sabo

Colon Biopsy Classification Using Crypt Architecture, in the 5th International Workshop in Machine Learning and Medical Imaging MLMI, 181-188, Boston, September 2014.

An extended version will be submitted to IEEE Journal of Biomedical and Health Informatics