Abstract
Attention-based convolutional networks have attracted great interest in recent years and achieved great success in improving representation capability of networks. However, most attention mechanisms are complicated and implemented by introducing a large number of extra parameters. In this study, we proposed a lightweight attention-based convolutional network (ConvNet-CA) that has a low computation complexity yet a high performance for brain disease detection. ConvNet-CA weights the importance of different channels in features maps and pays more attention to important channels by introducing an efficient channel attention mechanism. We evaluated ConvNet-CA on a publicly accessible benchmark dataset: Whole Brain Atlas. The brain diseases involved in this study are stroke, neoplastic disease, degenerative disease, and infectious disease. The experimental results showed that ConvNet-CA achieved highly competitive performance over state-of-the-art methods on distinguishing different types of brain diseases, with an overall multi-class classification accuracy of 94.88 +/- 3.64%.