PRA-UNet3+:全尺度跳跃连接CT肝脏图像分割模型  

PRA-UNet3+: Full-scale Connected CT Liver Image Segmentation Model

作　　者：钟经纬 ZHONG Jing-wei(School of Artificial Intelligence and Computer Science,Jiangnan University,Wuxi 214122,China)

机构地区：[1]江南大学人工智能与计算机学院,江苏无锡214122

出　　处：《软件导刊》2023年第2期15-20,共6页Software Guide

基　　金：国家自然科学基金项目(61876072,61902153,62072243,61772273);国家重点研发计划项目(2017YFC1601800);中国博士后科学基金项目(2018T110441);江苏省六大人才高峰项目(XYDXX-012)。

摘　　要：器官损伤死亡率高,严重威胁着人类的生命安全。人体内脏形态多样,解剖结构复杂,因此器官图像的准确分割有助于医生进行诊断。医学图像对高精度分割模型的需求很大,然而,大多数医学图像分割模型都是直接从一般的图像分割模型迁移过来的,常常忽略了浅层特征信息以及边界的重要性。为解决该问题,提出使用注意力门和点采样方法获得高质量分割边界的图像分割模型。在常用的肝脏医学图像数据集CHAOS上对该模型进行评估,平均Dice达到0.946 7,平均IoU达到0.962 3,平均F1 Score达到0.935 1,证明该模型可同时学习图像细节特征和全局结构特征,能更好地对肝脏图像进行分割。Organ lesions have a high mortality rate and seriously threaten the safety of human life. The internal organs of human body are diverse in form and complex in anatomical structure,accurate segmentation of the organ assists the doctor in making the diagnosis. High precision segmentation model is required for medical image. However, most segmentation models are directly transferred from the general image segmentation model. These models often ignore the importance of shallow feature information and boundaries. In order to solve this problem, attention mechanism and point sampling technique are proposed to obtain high quality segmentation boundary. The model was evaluated on CHAOS, a commonly used liver medical image dataset, and the average Dice was 0.946 7, the average IoU was 0.962 3, and the average F1 Score was 0.9351. It is proved that this model can learn both the detail features and the global structure features of the image, and can perform better segmentation of the liver image.

关 键 词：医学图像分割 U-Net 注意力门 点采样技术 

分 类 号：TP391.41[自动化与计算机技术—计算机应用技术]