Dave J, Gingold E, Yorkston J, Bercha I, Goldman L, Walz-Flannigan A, Willis C
Thomas Jefferson UniversityHospital, Philadelphia, PA.
Carestream Health, Inc, Penfield, NY.
Med Phys. 2012 Jun;39(6Part5):3648-3649. doi: 10.1118/1.4734820.
Anomalous pixels may be defined as those pixels whose exposure response relationship is deviant from the typical, expected or calibrated response. A group of anomalous pixels may Result in visible correlated artifacts. Here we demonstrate an approach to identify anomalous pixels and correlated artifacts using flat-field images.
Using manufacturer specific calibration geometry, sets of four flat-field images per detector were obtained with varying input air kerma values (0.5 to 160 μGy) from 9 digital detectors at 6 institutions. Images obtained before and after calibration, with both proper and improper gain maps and structured artifacts were additionally acquired with some detectors. Image analysis methodology under consideration by AAPM Task Group 150 was used.After eliminating 10mm borders, images were divided into square regions (100mm ). Anomalous pixels were identified as pixels within each region with valuesabove or below ±3 standard deviations (SD) relative to the mean value of the region. If these pixels were identified in all four images comprising a set, then they were reported as anomalous. Line artifacts were identified as rows and columns with cumulative profile values that were above or below ±3 SD with respect to the mean value of neighboring profiles in the set of four flat-field mages. Results were verified with visual inspection of the images.
For four sets of images, the algorithm did not identify any anomalous pixels, and none were spotted on visible inspection as well, while for five sets of images the identified anomalous pixels matched visual inspection results. Anomalous pixel detection failed in regions with an unusually large number of defects and structured noise, since those regions exhibited relatively large SD. Line artifacts consistent with visual analysis were identified correctly when present.
A practical approach to identify anomalous pixels and correlated artifacts from flat-field images is demonstrated.
异常像素可定义为那些曝光响应关系偏离典型、预期或校准响应的像素。一组异常像素可能会导致可见的相关伪影。在此，我们展示一种使用平场图像识别异常像素和相关伪影的方法。
利用制造商特定的校准几何结构，在6家机构从9个数字探测器获取了每组包含四个平场图像的数据集，输入空气比释动能值范围为0.5至160μGy 。部分探测器还额外获取了校准前后的图像，包括正确和不正确的增益图以及结构化伪影。采用了美国医学物理师协会任务组150正在考虑的图像分析方法。去除10mm边界后，将图像划分为正方形区域（100mm ）。异常像素被识别为每个区域内相对于该区域平均值其值高于或低于±3标准差（SD）的像素。如果在构成一组的所有四张图像中都识别出这些像素，则将它们报告为异常像素。线伪影被识别为在四张平场图像组中，其累积轮廓值相对于相邻轮廓平均值高于或低于±3 SD的行和列。通过对图像的目视检查验证结果。
对于四组图像，该算法未识别出任何异常像素，目视检查也未发现异常像素；而对于五组图像，识别出的异常像素与目视检查结果相符。在存在大量缺陷和结构化噪声的区域，异常像素检测失败，因为这些区域的标准差相对较大。当存在与视觉分析一致的线伪影时，能够正确识别。
展示了一种从平场图像识别异常像素和相关伪影的实用方法。
