DANet: Attention-based Dilated Network for Medical Image Segmentation

Abstract Colorectal cancer is a significant public health problem worldwide, and early detection is very necessary. However, the miss rate during routine colonoscopy examinations is very high, leading to undiagnosed polyps that can develop into colorectal cancer. Therefore, we proposed a novel architecture called Attention-based Dilated Network (DANet) for automatic polyp segmentation using convolutional neural networks (CNN). DANet uses a pre-trained ResNet50 proposed by He, Kaiming, et al [8] as an encoder and a Dilated Attention Convolution (DAC) block in between the encoder and decoder to learn a more robust feature representation. We evaluated DANet’s performance on four datasets namely KvasirSEG , CVC-ClinicDB , ETIS-Larib PolypDB and KvasirInstrument datasets and found that it outperformed state-of-the-art methods in terms of standard segmentation metrics, such as Jaccard score, F1 score, recall, accuracy, and F2 score. The superiority of DANet’s performance can be attributed to the effective combination of the pre-trained ResNet50 proposed by He, Kaiming, et al., [8] and the DAC block, which enables the network to learn long-range dependencies and focus on informative regions in the input image. This feature enables the network to segment the polyps accurately. In conclusion, the proposed DANet architecture is a promising solution for automatic polyp segmentation in the medical domain. Automated polyp segmentation methods such as DANet have the potential to improve detection rates and reduce miss rates, facilitating early detection and prevention of colorectal cancer, which can have a significant impact on public health. Keywords DANet, Colonoscopy, DAC, ResNet50, CNN.