一种基于相关性特征融合的乳腺图像感兴趣区域检测方法

A region of interest detection method for mammography based on multi-cue feature integration

乳腺图像的感兴趣区域(region of interest,ROI)检测是计算机辅助诊断乳腺疾病的第一步,检测效果的提升对减小误诊率有重要的作用.传统方法往往提取单独的视觉特征来描述乳腺图像,通过分类的方法找出包含肿块的区域.然而由于乳腺图像内容丰富结构复杂,使用单一的底层视觉容易忽视特征间的相互联系.提出基于相关性特征融合的乳腺图像ROI检测框架(multi-cue integration detection,MCID),通过引入乳腺图像的相关性特征,并与乳腺图像局部视觉特征相融合,辅助乳腺图像ROI的检测,以提高检测准确性.乳腺图像ROI检测实验表明,MCID可提高肿块检测的准确性.

ROI（region of interest）detection for mammography is the first step of CAD（computer-aided diagnosis）for breast disease.Mass is the region what we are interested in mammography,so the purpose of this paper is to improve mass detection because that will help to reduce the rate of misdiagnosis.Lots of work has focused on this field these years.Traditional methods only extract single visual feature for classification to detect the mass.However,mammography is complex.Employing only single visual feature will ignore the connection between images and features.That＇s why context information of mammography should be taken into consideration.Besides,feature integration is another important method to make up such deficiency which has been implemented in different ways and achieved good results recent years.But few work has considered both context information of mammography and feature integration method to improve ROI detection.As a result,in this paper a ROI detection method named MCID（multi-cue integration detection）is proposed based on multi-cue feature integration.The core idea ofthis method is integrating correlation features of context and several visual features together.First of all,both global and local correlation features are extracted by using the proposed multi-cue PLSA model（MCPLSA）,which is an improvement of PLSA,and then a classification model named CFIBM（correlation feature integration based boosting model）is proposed in order to improve the accuracy of mammography ROI detection.This model is made up of several logistic regression classifiers whose input has taken context information into consideration and output is the probability of being predicted as each label（mass or not）.At last,the final result is voted by each classifier.Experimental results on DDSM（digital database for screening mammography）show that MCID performs better than existing methods in two aspects,one is improving the accuracy of mass detection and the other is shortening the time needed.