Nuclei Detection Based on Single-point Labels

Abstract

Whole-slide image analysis is a long-lasting and laborious process. There are many ways of automatic analysis for histological images. The nuclei detection and classification is one of the most common and medically meaningful medical information-rich methods. However, sometimes the goal of nuclei detection is not to provide detailed information for the medical professionals but to be used for further aggregation. In such cases, nuclei segmentation exceeds requirements and takes extra resources during the data annotation. Keeping this in mind we optimized the existing state-of-art method for nuclei segmentation and classification to work with nucleus centers as input data. Combined with novel optimization technique and neural network activation function it resulted in the algorithm with has improved performance, easier training process and uses input data that is faster to produce.