LMCNet: a lightweight and efficient model for multi-classification of breast cancer images

Breast cancer is currently one of the most common cancers among women worldwide. Computer-aided diagnosis (CAD) is crucial in reducing diagnostic costs and improving subsequent cancer treatment. Convolutional neural networks are widely used in CAD, but most existing models are high-complexity and multi-parameter architectures, which limits their practical application. To address this issue, this paper proposes LMCNet, a lightweight breast cancer multi-classification model. LMCNet reduces the model complexity by refactoring the basic units of ShuffleNetV2, introducing PAConv parallel convolution layers and combining depthwise separable convolution. In order to further improve the feature extraction ability of the model, an SAE attention module containing channel and spatial dimensions is introduced. The proposed model achieves high recognition accuracy with low resource consumption. Our method requires only 275.69 K parameters and 51.19 M FLOPs. Compared to ShuffleNetV2, LMCNet reduces parameters by approximately 78.63% and FLOPs by 66.18%. Compared with other classical models, the proposed model achieves the best recognition efficiency and the fastest inference speed.