Abstract
Liver cancer is the leading cause of mortality in the world. Over the years, researchers have spent much effort in developing computer-aided techniques to improve clinicians' diagnosis efficiency and precision, aiming at helping patients with liver cancer to take treatment early. Recently, attention mechanisms can enhance the representational power of convolutional neural networks (CNNs), which have been widely used in medical image analysis. In this paper, we propose a novel architectural unit, local cross-channel recalibration (LCR) module, dynamically adjusting the relative importance of intermediate feature maps by considering the roles of different global context features and building the local dependencies between channels. LCR first extracts different global context features and integrates them by global context integration operator, then estimates per channel attention weight with a local cross-channel interaction manner. We combine the LCR module with the residual block to form a Residual-LCR module and construct a deep neural network termed local cross-channel recalibration network (LCRNet) based on a stack of Residual-LCR modules to recognize live cancer atomically based on CT images. Furthermore, This paper collects a clinical CT image dataset of liver cancer, AMU-CT, to verify the effectiveness of LCRNet, which will be publicly available. The experiments on the AMU-CT dataset and public SD-OCT dataset demonstrate our LCRNet significantly outperforms state-of-the-art attention-based CNNs. Specifically, our LCRNet improves accuracy by over 11% than ECANet on the AMU-CT dataset. SUPPLEMENTARY INFORMATION: The online version contains supplementary material available at 10.1007/s13755-023-00263-6.