Nuclei probability and centroid map network for nuclei instance segmentation in histology images

• Published in: Neural computing & applications 2023-07, Vol.35 (21), p.15447-15460
• Main Authors: Rashid, Syed Nauyan, Fraz, Muhammad Moazam
• Format: Article
• Language: English
• Subjects: Artificial Intelligence; Centroids; Coders; Computational Biology/Bioinformatics; Computational Science and Engineering; Computer Science; Data Mining and Knowledge Discovery; Digital imaging; Histology; Image Processing and Computer Vision; Image segmentation; Instance segmentation; Lightweight; Masks; Medical prognosis; Nuclei; Original Article; Pathology; Probability and Statistics in Computer Science; Survival analysis; Workflow
• ISSN: 0941-0643; 1433-3058
• DOI: 10.1007/s00521-023-08503-2

Nuclei instance segmentation is an integral step in digital pathology workflow as it is a prerequisite for most downstream tasks such as patient survival analysis, precision medicine, and cancer prognosis. There exist many challenges such as quality of labeled data, staining variation among tissue slides, high variation among multi-organ & multi-center digital slides and overlapping nuclei that are difficult to separate. Therefore, it is important to have an automatic and robust nuclei instance segmentation model that saves the time of pathologists by delineating accurate nuclei instances. To this end, we develop a nuclei instance segmentation pipeline that estimates distance transform and nuclear masks using an encoder–decoder-based CNN model. These estimated distance transform and nuclear masks are then utilized to delineate accurate nuclei boundaries from overlapping nuclei. We demonstrate that our proposed NC-Net model is lightweight and produces state-of-the-art results on the three recently published largest nuclei instance segmentation datasets to date. Additionally, our proposed NC-Net model is faster and utilizes a fewer number of parameters for learning as compared to other top-performing nuclei instance segmentation models. The purpose of developing a lightweight and state-of-the-art model is to provide capacity building to digital pathology workflows by reducing inference times and delineating accurate nuclear instances. The implementation details and the trained models are made available at this https://github.com/nauyan/NC-Net .