IDC-Net: Breast cancer classification network based on BI-RADS 4

•We construct a novel lightweight network named IDCNet to realize 3-classification tasks, BI-RADS 4 sub-categories 4a, 4b, and 4c.•The IDCNet has two paths, CNN and CapsNet, which combines the advantages of both. CNN is used to extract rich and deep local semantic features, and CapsNet is used to extract the position and posture information of the image.•For the CNN path, we design a network named ID-Net, which is composed of two sub-paths: one is constructed by superimposed ID block to extract rich semantic information from bad artifact images, and the other is constructed by superimposed DD block to better extract boundary information.•To realize the lightweight of the network, we use a special operation when fusing the information extracted by the two paths, which can control the growth of parameters caused by the combination of CNN and CapsNet. In the diagnosis of breast cancer, the 3 sub-categories 4a-4c of BI-RADS 4 are of great significance to doctors. However, low resolution of ultrasound image and high similarity between different category images pose great challenges to this task, which requires the network to be more capable of extracting image features. Therefore, in response to the efficient classification of BI-RADS 4a-4c in breast ultrasound images, we developed a lightweight classification network IDCNet, a neural network model combining the advantages of convolutional neural network(CNN) and CapsNet. In this model: Firstly, we proposed ID-Net based on CNN architecture and mainly constructed by ID block and DD block, which ensure the ID-Net deep and wide enough to extract sufficient local semantic information of image, and at the same time being lightweight. Secondly, we use the CapsNet to learn the position and posture information between the global features of the image, which makes up for the defects of CNN. Finally, two parallel paths of IDCNet and CapsNet are fused to enhance IDCNet's capability of feature extraction. To verify our method, experiments have been conducted on the breast ultrasound dataset of Yunnan cancer hospital and two public datasets. The classification results of our method have been compared with those obtained by five existing approaches. The experimental results show that the proposed method IDCNet has the highest Accuracy (98.54 %), Precision (98.54 %) and F1 score (98.54 %).