A multiscale attention network for grading of invasive ductal cancer

Invasive ductal carcinoma (IDC) represents the most prevalent form of breast cancer. Automatic and precise IDC grading is essential for the clinical assessment of tumor status. However, current methodologies struggle with accurate grading due to the intricate spatial structure within IDC tumor regions, coupled with significant similarities between different classes of images and high variability within the same class, posing a substantial challenge for IDC grading. To tackle these challenges, we propose a novel multi-scale attention model named MS-HiFuse. This model integrates the multi-scale convolution and the multi-head attention mechanism into the local feature block of the HiFuse network, creating an enhanced multi-scale local feature block that more effectively captures the nuanced features of the tumor area and facilitates the learning of fine-grained feature representations. Furthermore, to overcome the limited number of original data and to ensure data quality, we augment the dataset through operations such as maximum rectangle cropping, subdivision of plots, and image flipping for each sample. Comparison experiments demonstrate that our proposed MSHiFuse network attains an AUC of 80.84%, outperforming both the current competing networks and the original model. Hence, the application of the MS-HiFuse network proposed herein holds significant promise for the grading of IDC pathology images.