Krupinski E A, Nishikawa R M
Department of Radiology, University of Arizona, Tucson 85724, USA.
Med Phys. 1997 Jan;24(1):17-23. doi: 10.1118/1.597941.
The purpose of this study was to compare identifications of microcalcification clusters on mammograms by a computerized detection scheme and by human observers having their eye position recorded. Eighty digitized mammograms (half with a subtle microcalcification cluster) were analyzed by a computerized detection scheme and then were read from laser-printed films by six mammographers while eye position was recorded. The computer had 83% true positives with a false-positive rate of 0.5 per image. The true positives of the radiologists ranged from 78% to 90%, with false-positive rates ranging from 0.03 to 0.20. Locations of true and false positives identified by computer and by the human were compared. All but 5% of the true clusters were identified by either the computer, human, or by both. Here 10% of the clusters were detected by only the computer, and 11% were missed by the computer but detected by at least one radiologist. False positives were of three types: identified by computer only, by the human reader only, or by both. Eye-position data indicated significant differences in dwell time between both true-positive and false-positive locations reported by the radiologist versus the computer detections. A follow-up analysis indicated that microcalcification clusters and false positives were judged to have more identifiable characteristics of true calcifications and were associated with longer gaze durations than those with fewer microcalcification characteristics. In general, the computer was able to detect clusters judged to have few or no features that the radiologists were not able to detect. Comparison of computer versus human identification of microcalcification clusters may be useful for improving computerized detection schemes to serve as clinical aids to mammographers, and for understanding what image features lead to false-positive decisions for both the computer and the human reader.
本研究的目的是比较计算机检测方案与记录了眼动位置的人类观察者对乳腺钼靶片上微钙化簇的识别情况。采用计算机检测方案对80幅数字化乳腺钼靶片（其中一半含有细微的微钙化簇）进行分析，然后由6位乳腺造影医师在记录眼动位置的同时阅读激光打印片。计算机的真阳性率为83%，每张图像的假阳性率为0.5。放射科医师的真阳性率在78%至90%之间，假阳性率在0.03至0.20之间。比较了计算机和人类识别出的真阳性和假阳性的位置。除5%的真簇外，其余所有真簇均由计算机、人类或两者共同识别。其中，10%的簇仅由计算机检测到，11%的簇计算机未检测到但至少有一位放射科医师检测到。假阳性有三种类型：仅由计算机识别、仅由人类阅片者识别或两者共同识别。眼动位置数据表明，放射科医师报告的真阳性和假阳性位置与计算机检测结果相比，注视时间存在显著差异。后续分析表明，与微钙化特征较少的情况相比，微钙化簇和假阳性被判定具有更多真正钙化的可识别特征，并且与更长的注视持续时间相关。总体而言，计算机能够检测出放射科医师无法检测到的、被判定为特征很少或没有特征的簇。比较计算机与人类对微钙化簇的识别，可能有助于改进计算机检测方案，使其成为乳腺造影医师的临床辅助工具，并有助于理解哪些图像特征会导致计算机和人类阅片者做出假阳性判断。
