基于视觉感知信息的乳腺钼靶肿块检测分析与自动提取  

Applying Visual Perception Information for Detection Analysis and Automatic Extraction of Breast Mass in Mammograms

作　　者：柯尔挺[1] 厉力华[1] 刘伟 徐伟栋[1] 张娟[3] 张凌男[3] ZHENG Bin 

机构地区：[1]杭州电子科技大学生命信息与仪器工程学院,杭州310018 [2]Department of Radiology in University of Pittsburgh,PA 15213,USA [3]浙江省肿瘤医院放射科,杭州310022

出　　处：《中国生物医学工程学报》2014年第1期28-36,共9页Chinese Journal of Biomedical Engineering

基　　金：国家重点基础研究发展计划资助(2013CB329502);国家自然科学基金(61001215、61271063)

摘　　要：医生的视觉感知信息与影像诊断联系紧密,如何有效利用该信息以提高医学影像辅助诊断中的决策准确性,是一个具有前沿性和实际临床价值的研究课题。针对医生临床诊断时其视觉感知行为的分析和利用,探讨医生读片时单纯的视觉注视信息在多大程度上反映肿块位置(可检测性),以及如何利用视觉注视信息提取病灶。首先,用眼动仪采集医生读片时的注视点序列,每个注视点包括该点在钼靶影像中的相对位置、注视点停留时间和瞳孔直径等3个视觉特征,然后基于这些特征对注视点序列进行聚类分析,根据关注度评价找出医生浏览影像时的若干"关注点"位置,并对比分析其与肿块位置的关系,以评价"命中率";以关注点为引导,利用区域生长和水平集方法对肿块病灶进行提取。利用DDSM数据库和浙江省肿瘤医院数据库的75张钼靶影像进行初步实验。将关注点数限制在4个以内时,肿块病灶命中率为58.49%,同时所有命中肿块中被完整提取的占70.97%。结果表明,医生视觉注意信息对肿块位置的反应有一定作用,有助于理解感知反馈提高诊断精度的内在机理。Clinical diagnosis according to medical images is a process of radiologists＇ visual perception and decision-making. The radiologists＇ visual perception information is intimately associated with diagnosis. How to effectively use visual perception information to improve the decision-making accuracy in computer-aided diagnosis is a research subject which is full of scientific significance and clinical value. This paper conducted researches on the analysis and the use of visual perception behavior during diagnosis to explore two issues ： one is how well the single perceptual information during diagnosis can reflect masses position; the other is how to use the perceptual information for extracting masses. In the paper, the research method includes two steps. Firstly, radiologists＇ fixation point sequence, in which every point includes fixation point location in mammogram, duration time and pupil diameters, was recorded by an eye-tracker during reading and then clustered to achieve some radiologists＇ concerns according to the three visual features. Shooting average was calculated by analyzing the positional relationship between concerns and masses in the same mammogram. Secondly, regarding concerns as seeds, the SBRG （seeds-based region growing） approach and the multi-scale mass segmentation approach were applied to buckle breast masses from mammograms. The result of applying the proposed method to 75 mammograms from both DDSM and Zhejiang Cancer Hospital showed that it could achieve shooting average of 55.49% when the limitation of concerns number was 4, and the full-extraction rate for the shot masses was 70.97%. It is revealed that the perceptual information is helpful to reflect masses position as well as to understand the inner mechanism of perceptual feedback.

关 键 词：视觉感知 图像分割 眼动仪 聚类分析 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TP751.1[自动化与计算机技术—计算机科学与技术]