LMBNet: Lightweight Multiple Branch Network for Recognition of HER2 Expression Levels

In recent years, more research methods for HER2 automatic evaluation have been presented. However, these methods are complex and expensive. In this paper, we present a lightweight, highly modular network architecture for HER2 classification. The presented network consists of multiple LMB modules, and the LMB module includes two parts: the downsampling and feature extraction. Both the two parts are multi-branch processing structures to accommodate the complexity of HER2 image features. Especially, in the feature extraction part, dilated convolution is used to enrich the feature extraction scale and reduce the training parameters. Moreover, grouped convolution is used to process the transferred feature maps inside and between modules to reduce the reuse of feature maps. Then, the channel shuffle is used to avoid the isolation of feature information. Compared with other HER2 classification methods, the presented model has lower computational cost while ensuring HER2 automatic evaluation performance.