Implementing Attention Mechanism in Convolutional Neural Network to Improve Performance of MRI Image Classification of Nasopharyngeal Cancer

Cancer is one of the main diseases that threaten human death, and nasopharyngeal cancer also shows a high mortality rate. The early diagnosis is particularly important in the proper treatment of cancers. Computer-aided diagnosis has been widely used in the medical field. To harness the artificial intelligence in medical imaging, we implement two types of attention mechanism in the popular convolutional neural network ResNet50 to aid classification and diagnosis of the medical images of nasopharyngeal cancer. Compared with basic ResNet50 architecture, both “Convolutional Block Attention Module (CBAM)” and “Dual Attention Network (DANet)” gain the improved classification performance. Our results show that the implementing location affects the results. We compare six types of implementing ways, named as CBAM-A, CBAM-B, DANet-A, DANet-B, Fusion-A and Fusion-B. Among six models, DANet-B implemented network achieves the 96.5% accuracy, 96.8% precision, 96.5 % recall and 96.4 % F1-score, showing significant improvement compared with the basic ResNet50 with values of 54.4% accuracy, 60.5% precision, 54.4% recall and 50.6% F1-score, respectively. The results show that proper implementing attention mechanism can improve the classification performance and may be developed as an auxiliary diagnosis approach for the Nasopharyngeal Cancer.