A multi-input lightweight convolutional neural network for breast cancer detection considering infrared thermography

Abstract

Although deep convolutional neural network (CNN) has been widely used in the breast cancer detection based on thermal imaging technology, this scenario still did not receive enough attention in the mobile devices with limited resource. In addition, there still exists challenge on how to assist front view thermal imaging by side one during breast cancer detection. This study proposes a multi-input lightweight CNN named Multi-light Net in order to achieve more accurate early detection for breast cancer, which combines the thermal image from multiple perspectives with the lightweight CNN on the basis of model performance and scale. In addition, a new weighted label smoothing regularization (WLSR) is proposed for the Multi-light Net with the purpose of increasing the network’s generalization ability and classification accuracy. The experimental results demonstrate that the proposed approach by combining front view with side view can achieve more significant results than the common one using only front view during breast cancer detection, and the proposed Multi-light Net also exhibits an excellent performance with respect to the currently popular lightweight CNN. Furthermore, the proposed WLSR loss function can also lead to both faster convergence rate and more stable training process during network training and ultimately higher diagnostic accuracy for breast cancer.