Improved Semantic Segmentation in Medical Imaging Using U-Net and Attention Mechanisms

Abstract

Image segmentation poses a significant challenge in the field of medical image analysis aiming to extract valuable information and enhance clinical diagnosis accuracy. The main aim of this paper is to explore the use of the baseline U-Net architecture and another model combine between baseline U-Net with attention mechanisms called OrganFocusUNet model in organ semantic segmentation and differentiation in laparoscopic hysterectomy. These models involve leveraging the UD Ureter-Uterine Artery-Nerve dataset, which are a comprehensive collection from laparoscopic surgeries, accompanied by corresponding multiclass masks and capable of pixel-wise detection and differentiation of three key organs: ureter, uterine artery, and nerves, with a specific emphasis on accurately distinguishing the ureter from the other organs. The experiments showed that the baseline U-Net model on the augmented dataset have mean IoU score of 79.04%, while the proposed OrganFocusUNet model achieved on the augmented dataset a mean IoU score of 79.52%, indicating its effectiveness in accurately distinguishing critical organs.