Deep learning-powered multiple-object segmentation for computer-aided diagnosis

Recently, driven by hardware devices and deep learning technologies, computer-aided diagnosis systems have been widely applied, such as cancer diagnosis and early screening of autistic children. Many studies have reported extracting tumor regions from whole-slide images (WSI) in cancer diagnosis tasks, namely, image segmentation. However, doctors must re-analyze the ROI in the tumor area for some challenging diseases. Efficient segmentation algorithms are the key parts of perfecting machine diagnostic assistance systems. This paper presents a novel WSI segmentation framework (called UFINet), aiming to segment the tumor region on the liver tissue image and re-segment the region of interest in the tumor. The proposed algorithm provides a solution for applying medical human-computer interaction systems. The proposed framework was trained and tested on the liver tissue dataset and achieved a Dice of 66% on 86 WSIs. Experiments prove that the proposed UFIN et achieves top performance and meets the clinical requirements, providing an effective method for developing computer-aided diagnosis systems.