RBCA-Net: Reverse Boundary Channel Attention Network for Kidney Tumor Segmentation in CT images

Recently, as the importance of early diagnosis and treatment of cancer has increased, many studies have been introduced to analyze medical images using deep learning. In medical image analysis task, the lesions segmentation methods uses a Fully Convolutional Network (FCN) architecture such as U-Net to predict the lesion area and play an auxiliary role in medical care. So many researchers are working on improving the performance of architectures. But, there are some challenges in that data is imbalanced and the size and shape of lesions are irregular. To solve these problems, we improved the segmentation performance by using a two-stage cascaded method. In stage 1, coarse region of interest (RoI) was extracted using ResUNet, In stage 2, we use Atrous Spatial Pyramid Pooling (ASPP) to extract features to contain a lot of spatial information using various receptive fields from a pretrained DenseNet-161 backbone. In addition, we introduce the RBCA module that combines Reverse, Boundary, and Channel Attention to capture various sizes and shapes of lesions. The performance of the proposed model shows high performance compared to various architectures using the KiTS19 dataset including kidney and tumor.