An integrated convolutional neural network with attention guidance for improved performance of medical image classification

Abstract

Today, it becomes essential to develop computer vision algorithms that are both highly effective and cost-effective for supporting physicians' decisions. Convolutional Neural Network (CNN) is a deep learning architecture that enables learning relevant imaging features by simultaneously optimizing feature extraction and classification phases and has a high potential to meet this need. On the other hand, the lack of low- and high-level local details in a CNN is an issue that can reduce the task performance and prevent the network from focusing on the region of interest. To tackle this issue, we propose an attention-guided CNN architecture, which combines three lightweight encoders (the ensembled encoder) at the feature level to consolidate the feature maps with local details in this study. The proposed model is validated on the publicly available data sets for two commonly studied classification tasks, i.e., the brain tumor and COVID-19 disease classification. Performance improvements of 2.21% and 1.32%, respectively, achieved for brain tumor and COVID-19 classification tasks confirm our assumption that combining encoders recovers local details missed in a deeper encoder. In addition, the attention mechanism used after the ensembled encoder further improves performance by 2.29% for the brain tumor and 6.13% for the COVID-19 classification tasks. Besides that, our ensembled encoder with the attention mechanism enhances the focus on the region of interest by 4.4% in terms of the IoU score. Competitive performance scores accomplished for each classification task against state-of-the-art methods indicate that the proposed model can be an effective tool for medical image classification.