Development of Similar Image Subtraction Method for Digital Chest Radiographs

我们开发了一种相似图像减影方法的计算机方案，该方案提供了从不同患者的相似胸片中获得的减影图像。在本研究中，待诊断的图像称为“目标图像”。通过岩手县肺癌筛查获得的8776张正常胸片（5352张女性胸片，3424张男性胸片）建立了一个大型图像数据库。所有图像均取自计算机x线摄影（CR）系统。从一个由正常胸片组成的大型数据库中搜索目标图像的相似图像。首先，根据目标图像与正常图像之间肺大小的相似性选择相似图像的初始候选图像（数据库中10%）。其次，利用目标图像与候选图像之间的相关值，从初始候选图像中进一步选择相似图像；最后，从目标图像中选取不同患者的相似图像进行相减，生成相减图像。基于非线性图像翘曲技术，采用全局匹配和局部匹配相结合的方法进行图像减法。从不同患者的相似胸片中获得的减影图像的质量根据目标图像与相似图像之间的误配而产生的伪影进行主观评估。从数据库中选择的相似图像与目标胸片非常相似。虽然从不同患者获得的减影图像包含一些伪影，但其图像质量对于异常病变的检测是可以接受的。从相似图像中获得的约40%的减影图像被认为在临床上有助于增强胸片上的异常。

We have developed a computerized scheme of similar image subtraction method which provides subtraction images obtained from similar chest radiographs of different patients. In this study, an image to be diagnosed is called "target image".1.A large image database was established with 8776 normal chest radiographs(5352 female images and 3424 male images) obtained from a lung cancer screening in Iwate prefecture. All images were taken from a computed radiography(CR) system.2.Similar images of a target image were searched from a large database consist of normal chest radiographs. First, initial candidates of the similar image(10% in the database) were selected by the similarity of the lung size between the target image and normal images. Second, further similar images were chosen from the initial candidates by using their correlation values between the target image and candidates.3.Finally, subtraction images were produced by subtracting the selected similar images of different patients from the target image. The subtraction was performed with a global matching and local matching based on a nonlinear image warping technique.4.The quality of the subtraction images obtained from similar chest radiographs of different patients were evaluated subjectively in terms of artifacts due to misregistration between a target image and a similar image.5.Similar images selected from the database were remarkably resembled to the target chest radiographs. Although the subtracted images obtained from different patients included some artifacts, their image qualities were acceptable for detection of abnormal lesions. Approximately, 40% of the subtraction images obtained from the similar images were considered clinically useful to enhance abnormalities in chest radiographs.