An Automatic Segmentation of Gland Nuclei in Gastric Cancer Based on Local and Contextual Information

摘要

Analysis of tubular glands plays an important role for gastric cancer diagnosis, grading, and prognosis; however, gland quantification is a highly subjective task, prone to error. Objective identification of glans might help clinicians for analysis and treatment planning. The visual characteristics of such glands suggest that information from nuclei and their context would be useful to characterize them. In this paper we present a new approach for segmentation of gland nuclei based on nuclear local and contextual (neighborhood) information. A Gradient-Boosted-Regression-Trees classifier is trained to distinguish between gland-nuclei and non-gland-nuclei. Validation was carried out using a dataset containing 45702 annotated nuclei from 90 1024 × 1024 fields of view extracted from gastric cancer whole slide images. A Deep Learning model was trained as a baseline. Results showed an accuracy and f-score 5.4% and 23.6% higher, respectively, with the presented framework than with the Deep Learning approach.