基于注意力机制多任务的肺结节癌变风险判断  

Risk Assessment of Lung Nodule Canceration Based on Attention Mechanism and Multitask

作　　者：王广涵 程远志 史操 许灿辉 WANG Guang-Han;CHENG Yuan-Zhi;SHI Cao;XU Can-Hui(School of Information Science and Technology,Qingdao University of Science and Technology,Qingdao 266061,China)

机构地区：[1]青岛科技大学信息科学技术学院,青岛266061

出　　处：《计算机系统应用》2022年第4期117-122,共6页Computer Systems & Applications

基　　金：国家自然科学基金(61806107,61973180,62002190)。

摘　　要：对于CT影像中检测出的肺部结节,需要自动判断其是否有癌变风险.不同于大多数现有的研究方法只区分结节良恶性,本文提出了一个基于注意力机制的多任务学习模型,将与结节良恶性相关的语义特征属性一并判断输出,通过判断9个结节特征(对比度、分叶征、毛刺征、球形度、边缘、纹理、钙化程度、大小以及恶性程度)的同时实现内在特征的共享,以达到提高各子任务性能的目的.选择视觉转换器(ViT)模型作为多任务共享特征提取层,整体模型采用动态加权平均方法来对各子任务的Loss函数进行优化.在LUNA16数据集上的实验表明,该学习框架可以提升肺结节癌变风险判断的性能,且同时对其他语义特征的判断也能提升结果的可解释性.For pulmonary nodules detected in computed tomography(CT) images, it is necessary to automatically determine whether they are at the risk of canceration. This study proposes a multitask learning model based on the attention mechanism. Different from most existing research methods which only distinguish between the benignity and malignancy of nodules, the proposed model also assesses and outputs the semantic features related to the benignity and malignancy of nodules. The assessment of nine nodule features(subtlety, lobulation, spiculation, sphericity, margin,texture, calcification, diameter, and malignancy) and the sharing of internal characteristics are conducted at the same time to improve the performance of each subtask. The vision transformer(ViT) model is selected as the multitask shared feature extraction layer, and the whole model uses the dynamic weighted average method to optimize the Loss function of each subtask. Experiments on the LUNA16 dataset show that the proposed learning framework can improve the risk assessment of pulmonary nodule canceration and that the assessment of other semantic features can also enhance the interpretability of the results.

关 键 词：肺结节 癌变 低剂量螺旋CT 多任务 注意力机制 计算机辅助诊断 医学影像处理 

分 类 号：R734.2[医药卫生—肿瘤] TP391.41[医药卫生—临床医学]