MDAA: multi-scale and dual-adaptive attention network for breast cancer classification

Attention mechanism is crucial in the auxiliary diagnosis of breast cancer. However, methods relying on a single attention mechanism may not always achieve satisfactory results. To address this, we proposed the multi-scale and dual-adaptive attention network (MDAA) for breast cancer pathological image classification. It is a novel hybrid model based on DenseNet and a multi-scale feature extraction module, which incorporates dual-adaptive attention and adaptive balance loss function. Initially, dense block and multi-scale block serve as the network backbone, facilitating feature reuse and enhancing expressiveness. Subsequently, the dual-adaptive attention block is introduced to capture richer features. Finally, an adaptive balancing loss is used to handle the class imbalance problem. Through experiments on two public datasets, it is demonstrated that MDAA exhibits higher performance and is superior to the existing methods. Its strong robustness and generalization make it suitable for breast cancer auxiliary diagnosis.