肺部CT图像多病种自动检测及分类  

Automatic detection and classification of multiple lung diseases in CT images

作　　者：吴优 张睿[1] 张文强[1] WU You;ZHANG Rui;ZHANG Wenqiang(School of Computer Science,Fudan University,Shanghai 200011,China)

机构地区：[1]复旦大学计算机科学技术学院,上海200011

出　　处：《应用科技》2022年第2期27-32,共6页Applied Science and Technology

基　　金：上海市2020科技创新行动计划项目(20511103102)。

摘　　要：针对医学影像预处理复杂、病灶位置分散检测困难、医院等实际应用场景的设备条件难以满足庞大影像数据量对设备的高性能要求等难点,本文采用卷积神经网络的方法训练双阶段模型,对肺结节、索条和动脉硬化钙化3种病灶进行检测,第一阶段目标是检测查全率,第二阶段目标是检测查准率。实验结果表明本文方法在设备有限时,检测时间约为DeepLung等3D模型的10%,检测准确性比YOLOv3等2D模型要高,在实际应用场景中具有较高的实用性。Because of the complexity of medical image pretreatment and difficulty in detecting the scattered focus of infection, actual situations such as hospitals are not able to provide high-performance equipment to deal with large volume of image data. To solve this problem, this paper adopts the method of convolution neural network to train double-stage models, and tests three focuses of lung nodules, rope and hardening of the arteries calcification lesions. The target of the first stage is to test the recall rate, and the target of the second stage is to test the precision rate.Experimental results show that when equipment is limited, the detection time of the proposed model is about 10% of that of 3D models such as DeepLung, and that the detection accuracy is higher than that of 2D models such as YOLOv3. The method is effective in practical application scenarios.

关 键 词：医学影像 肺部多病种检测 卷积神经网络 CT图像病灶检出 人工智能医疗 2D模型 3D模型 双阶段模型 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]