Multi-task Deep Learning Pipeline for Irregular Macrophage Segmentation fusing Enhanced Microscopy Images and Auxiliary Boundary Maps

Abstract

Abstract Background:Accurate segmentation of macrophages from microscope images can be beneficial for understanding viral infection and immune development stages. There are two particularly challenging aspects in this task: (1) over-segmentation of macrophage with pseudopodia; (2)under-segmentation for clustering, overlapping macrophages and unclear boundary macrophages. Method: This paper proposes a microscope image enhancement multi-task deep learning framework to achieve segmentation of macrophages with complex boundary conditions. The network initially utilizes PENet to enhance the quality of microscope image data, improving the overall image quality. Then, a multi-task U-Net architecture is employed to extract crucial feature information from masks, distance transforms, and heatmaps. By utilizing the cell segmentation achieved through masks, the distance transforms and heatmaps are used to further refine and capture the intricate boundary details of macrophages, including pseudopodia and other irregularities. Results: Despite the challenges posed by partially or entirely obscured cells, the network demonstrates robust segmentation capabilities for surface-visible cells, achieving an accuracy of 61.24%, a precision of 78.79%, and a recall of 87.93%, outperforming some other segmentation networks, including SOTA Cellpose. Through experiments, it is possible to achieve precise segmentation of irregular boundaries and narrow pseudopodia of macrophages in low-quality microscope images. Conclusions: Compared to current macrophage segmentation techniques, this network has two significant advantages: (1) supplementing rich boundary detail information to capture the microscopic features of macrophageswith elongated pseudopods; (2) enhancing underexposed cells due to limitations of microscopic imaging techniques and capturing their potential information.