Specialized recurrent U-Net architecture for immunohistochemistry image segmentation

Abstract

Microobject segmentation in biomedical images is a complex and time-consuming process, and current efforts are aimed at qualitatively separating the studied objects from the background in the image. Immunohistochemical images are used for breast cancer analysis. One of the types of images is estrogen samples. The specificity of immunohistochemical images requires a more thorough approach to segmentation, because standard algorithms or classical neural network models do not allow achieving the required accuracy. This article presents a new convolutional network architecture based on a recurrent U-network for segmentation of immunohistochemical images. The proposed architecture is specifically designed to process the unique features of immunohistochemical images. The principles of recurrent convolutional neural networks were used to develop the model and train it using open datasets. The architecture proposed in this work is based on the configuration of neural network layers, which allows to obtain the best result for a given type of image. In addition, a pipeline of continuous integration of the program code and the model was implemented to ensure continuous retraining in a cloud environment, which significantly simplified the workflow for researchers and engineers. Additionally, a mechanism for processing medical images and providing automation of the big data processing process using computer vision and deep learning was proposed. The proposed method achieved higher segmentation accuracy. The architecture allows for more accurate cell segmentation on immunohistochemical images, which is a crucial step in the diagnostic pipeline. High-quality cell segmentation increases the reliability of subsequent analyses and parameter extraction, thereby contributing to increased diagnostic accuracy.