In computer-aided diagnosis methods for breast cancer, deep learning has been shown to be an effective method to distinguish whether lesions are present in tissues. However, traditional methods only classify masses as benign or malignant, according to their presence or absence, without considering the contextual features between them and their adjacent tissues. Furthermore, for contrast-enhanced spectral mammography, the existing studies have only performed feature extraction on a single image per breast. In this paper, we propose a multi-input deep learning network for automatic breast cancer classification. Specifically, we simultaneously input four images of each breast with different feature information into the network. Then, we processed the feature maps in both horizontal and vertical directions, preserving the pixel-level contextual information within the neighborhood of the tumor during the pooling operation. Furthermore, we designed a novel loss function according to the information bottleneck theory to optimize our multi-input network and ensure that the common information in the multiple input images could be fully utilized. Our experiments on 488 images (256 benign and 232 malignant images) from 122 patients show that the method’s accuracy, precision, sensitivity, specificity, and f1-score values are 0.8806, 0.8803, 0.8810, 0.8801, and 0.8806, respectively. The qualitative, quantitative, and ablation experiment results show that our method significantly improves the accuracy of breast cancer classification and reduces the false positive rate of diagnosis. It can reduce misdiagnosis rates and unnecessary biopsies, helping doctors determine accurate clinical diagnoses of breast cancer from multiple CESM images.