Abstract
Abstract. Mucous glands is an important diagnostic element in digestive pathology. The first step of differential diagnosis of colon polyps in order to assess their malignant potential is gland segmentation. The process of mucous glands segmentation is challenging as the glands not only needed to be separated from a background but also individually identified to obtain reliable morphometric criteria for quantitative diagnostic methods. We propose a new convolutional neural network for mucous gland segmentation that takes into account glands’ contours and can be used for gland instance segmentation. Training and evaluation of the network was performed on a standard Warwick-QU dataset as well as on the collected PATH-DT-MSU dataset of histological images obtained from hematoxylin and eosin staining of paraffin sections of colon biopsy material collected by our Pathology department. The collected PATH-DT-MSU dataset will be available at http://imaging.cs.msu.ru/en/research/histology.
Highlights
A differential diagnosis criteria of colon polyps are not accurate, there is no quantitative criteria of basal dilation of the crypts and spread of the serration as well as no principles for determining the malignant potential of various benign colon epithelial neoplasms
We propose a new architecture of a convolutional neural network (CNN) for mucous glands segmentation (Fig. 1) based on UNet model (Ronneberger et al, 2015) which has proven its good efficiency for segmentation of biomedical images
The second dataset is PATH-DT-MSU dataset that was collected and annotated by our Department of Pathology and consists of 20 histological images obtained from hematoxylin and eosin staining of paraffin sections of colon biopsy material. 13 images are hyperplastic polyps (HP); 6 images are sessile serrated adenomas (SSA/P) and one image is normal colon mucous glands
Summary
A differential diagnosis criteria of colon polyps are not accurate, there is no quantitative criteria of basal dilation of the crypts and spread of the serration as well as no principles for determining the malignant potential of various benign colon epithelial neoplasms. Almost all CNN-based segmentation methods (Long et al, 2015), (Badrinarayanan et al, 2017), (Ronneberger et al, 2015) use the same idea of convolutional autoencoder (CAE) (Masci et al, 2011). With minor changes these CNN-based segmentation methods can be applied to histological images. All three pipelines a fused into one, and are followed with several convolution layers to predict the final instance segmentation map This approach leads to the state-of-the-art level of segmentation accuracy
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
