A lightweight pathological image segmentation framework based on heterogeneous cross-layer sampling
Hai Shi, Hailong Chen, Yangyang Zhang, Zhengxia Wang · 2023
Pathological image segmentation is an essential step in the early detection and diagnosis of various diseases. The features of high resolution, complex background structure, and different tissue shape and size of pathological images, bring great challenges to the segmentation algorithm of pathological images. Deep learning is proved to be an effective method for the segmentation of pathological images. However, there are still some difficulties such as poor capture ability of lesion characteristics and complex model parameters. To solve the above problems, this paper designs a lightweight segmentation network framework for pathological image, called HES-UNet. The framework employs a learning mode of heterogeneous cross-layer sampling and feature capture, which improves computational efficiency while guaranteeing the performance of image segmentation. The way of heterogeneous cross-layer sampling is used in the encoder and decoder to perform only two feature extractions at large and small scales, which greatly reduces the number of parameters of the model. And after drastic dimensional changes, an attention mechanism is adopted for a global feature and local feature capture to improve the learning ability of the network on pathological features and the performance of image segmentation algorithm. Finally, to verify the performance of the proposed segmentation framework, we conducted experiments on three different datasets. The experimental results show that the network outperforms the contrast algorithm on segmentation performance and efficiency to obtain more accurate segmentation results. Compared with the standard UNet3+, the number of parameters can be reduced to 1/39 of the original. It is indicated that our proposed segmentation framework not only has efficient segmentation performance, but also has good generalization ability.