Gland cell image segmentation method and system based on improved U-Net network

Abstract

The invention discloses a gland cell image segmentation method and system based on an improved U-Net network. According to the method, the convolution layers are combined to construct the dense connection block, and the U-Net is amplified by a multi-resolution analysis method, so that the problem of cell segmentation with different rules can be well solved. And a self-attention module, namely a position attention module and a channel attention module, is introduced at a decoder end, so that context information can be adaptively aggregated, and the feature representation of cell segmentation isimproved. The cell image is segmented by using the network, so that the segmentation effect is improved. The method is easy to implement, data preprocessing operation is simple, and better robustnessand accuracy are achieved.

Description

technical field [0001] The invention relates to the technical field of network pattern recognition and segmentation, in particular to a non-diagnostic purpose-based glandular cell image segmentation method and system based on a densely connected block U-Net network. Background technique [0002] Histopathological image analysis plays a crucial role for accurate tumor cell estimation and prognosis. The size, shape, and some other morphological manifestations of imaged structures such as nuclei and glands are highly correlated with the presence or severity of disease. Previously, these images were evaluated by pathologists who evaluated biopsy samples, but the manual process was subjective, labor-intensive, and time-consuming. Therefore, computational methods to quantitatively and objectively analyze histopathology images have been developed. Image segmentation, i.e. the extraction of cells, nuclei or glands from histopathological images, is a step preceding all analysis and...

AI Technical Summary

Problems solved by technology

Despite the use of skip connections capable of propagating information from shallow to deep layers, the network fails to capture more detailed information due to the standard encoder-decoder architecture, where the dimensionality of the data is reduced near the bottleneck, the first few Blocks learn low-level features of the data, while later blocks learn high-level features, and eventually, the encoder learns to map the data to lower dimensions (in the spatial sense)

The ever-increasing receptive tiny structures in the network depth require smaller receptive fields, and in the case of standard U-Net, even with skip connections, the smallest receptive field is limited by the first layer

Thus, incomplete architectures inherently limit their ability to capture finer details

Images