Medical Image Segmentation Using Attention-based Residual Double U-Net

Abstract

A common use case for image segmentation in medical-image-based diagnosis is to help clinicians to focus on a specific area of the disease. Manually inspecting polyps from colonoscopy for colorectal cancer or performing a biopsy on skin lesions for skin cancer are time-consuming, laborious, and complex procedures. Automatic medical image segmentation aims to expedite this diagnosis process. The accuracy of image segmentation has increased due to advancements in machine learning techniques and deep learning models. However, there is still room for improvement as there exist various challenges due to the large variation in the appearance of objects in different sizes with no distinct boundaries. To address these issues, we propose a novel-attention based residual Double U-Net architecture that improves on the currently existing skin lesion segmentation networks. We incorporate attention gates on the skip connections and residual connections in the convolutional blocks of Double U-Net, a state-of-the-art (sota) segmentation network. The attention gates allow the model to retain more relevant spatial information by suppressing irrelevant feature representation from the down-sampling path. At the same time, residual connections help to train deeper models by ensuring better gradient flow. We conducted experiments on three datasets: ISIC 2018 (skin lesion), CVC Clinic-DB (polyp), and the 2018 Data Science Bowl (nuclei) datasets and achieved Dice Coefficient (DSC) scores of 91.64%, 94.35% and 92.45% respectively. Further improvement can be achieved by simplifying the structure of our architecture in order to reduce the number of parameters.