Improved DenseNet with Convolutional Attention Module for Brain Tumor Segmentation

In this paper, a novel Convolutional Attention Module (CAM) is proposed and integrated into the Densely Connected Convolutional Network (DenseNet) for brain tumor segmentation. This module makes it possible for the DenseNet to suppress irrelevant regions in brain MR images while highlighting useful salient features. Upon this integration, we combine the selectivity of the CAM and the feature reusability of the DenseNet to resolve the tumor segmentation problem where meaningful and important feature information is less visible while noise and interference are serious. Our proposed model was evaluated in Multimodal Brain Tumor Image Segmentation (BRATS 2015) datasets. In the online assessment of the datasets, the Dice Similarity Coefficient (DSC) values for the complete tumor area, core tumor area, and enhanced tumor area were 0.84, 0.72, and 0.63, respectively, which are superior to the results from the traditional U-net and the DenseNet itself. Experimental results demonstrate that our proposed model enables effective segmentation of brain tumors and can assist medical practitioners in analyzing complicated medical images.