Phys. Med. Biol. zyxwvutsrq 39 (1994) 2273-2288. Pcinted in the UK Computer-aided diagnosis in mammography: classification of mass and normal tissue by texture analysis Arthur Petrosian, Heang-Ping Chan, Mark A Helvie, Mitchell M Goodsitt and Dorit D Adler Department of Radiology, University of Michigan. A m Arbor, MI. USA Received 24 January 1994, in final form 16 August 1994 Abstract. Computer-aided diagnosis schemes zyxwv are being developed to assist radiologists in mammographic interpretation. In this study, we investigated whether texture features could be used to distinguish belween mass and non-mass regions in clinical mmmomms. Forry-five regions of interest (ROB) containing true masses with various degrees of visibility and 135 ROB containing normal breast parenchyma were extracted manually zyxw from digitized mammograms as case samples. Spatiill-grey-level-dependence zyxwvu (scw) matrices of each ROI were calculated and eight texture features were calculated from the SGLD matrices. The camlation and clawdistance properties of extracted texture feaNm were analysed. Selected texture features were input into B modified decision-tree classification scheme. The performance of the classifier was evaluated for different feature combinations and orders of features on the wee. A classification accumcy of about 89% sensitivity and 76% specificity was obtained for ordered features, sum average, correlation, and energy. during the training procedute. With a leave-one-out method, the test result was about zyxw 16% sensitivity and zyx 64% specificity. The results of this preliminaq study demansrmte the feasibility of using texture information for classification of mass and normal b r a t tissue, which will be likely to zyx be useful for classifying me and false detections in computer-aided diagnosis programmes. 1. Introduction It is known that images of many target objects or lesions are characterized by unique textural and shape properties. Computerized pattern-recognition techniques have been applied to mammographic images (Hand et al 1979, Magnin et al 1986, Chan et al 1987, Fam et al 1988, Lai er al 1989, Caldwell et al 1990, Davies et zyxwv a1 1990, Yin er al 1991, Cheng er al 1993) as well as other types of medical image (Kruger er al 1974, G m a et al 1989, Cheng et al 1991, Goldberg et al 1992). Detection of mammographic abnormalities using morphological features has been reported (Hand et a1 1979, Chan et al 1989, Lai et al 1989, Davies et al 1990, Brzakovic er al 1990, Yin et al 1991, Mascio et al 1993). Investigators have also explored the extraction of image statistical and textural information and classification zyxwvu of normal and disease patterns with discriminant analysis or neural- network techniques (Kruger et al 1974, Magnin er al 1986, Garra et al 1989, Katsuragawa et al 1988. Caldwell et al 1990, Cheng et al 1991, Goldberg et al 1992, Dhawan et al 1993, Cheng et al 1993). More recently, it was reported that edge-gradient orientation in combination with Laws texture features could be used to effectively detect spiculated masses on mammograms (Kegelmeyer et al 1994). The methods utilizing texture features take advantage of the fact that computers are better than human observers in analysing second-order statistical features. 0031-9155/94/122273+16$19.50 0 1994 IOP Publishing Ltd 2273