A comparison of computer based classi®cation methodsapplied to the detection of microaneurysms in ophthalmic¯uorescein angiogramsAllan J. Framea,*, Peter E. Undrilla, Michael J. Creea, John A. Olsonb,Kenneth C. McHardyc, Peter F. Sharpa, John V. ForresterbaDepartment of Bio-medical Physics and Bio-engineering, University of Aberdeen, Aberdeen, UKbDepartment of Ophthalmology, University of Aberdeen, Aberdeen, UKcDiabetic Clinic, Aberdeen Royal In®rmary, Aberdeen, UKReceived 10 October 1997; accepted 11 May 1998AbstractWe compared the performance of three computer based classi®cation methods when applied to theproblem of detecting microaneurysms on digitised angiographic images of the retina. An automatedimage processing system segmented `candidate' objects (microaneurysms or spurious objects), andproduced a list of features on each candidate for use by the classi®ers . We compared an empiricallyderived rule based system with two automated methods , linear discriminant analysis and a learningvector quantiser arti®cial neural network, to classify the object s as microaneurysms or otherwise. ROCanalysis shows that the rule based system gave a higher performance than the other methods ( p = 0.92)although a much greater development time is required. # 1998 Elsevier Science Ltd. All rights reserved.Keywords: Neural networks; Linear discriminant analysis; Rule based system; Ophthalmology; Computer aided diag-nosis1. IntroductionDiabetic retinopathy (DR) is the ocular manifestation of the systemic disease, diabetesmellitus and is the most common cause of blindness in the UK working population [1]. Incurrent research studies, assessment of DR is made semi-quantitatively by comparing aphotograph of the patient's fundus with a standard set of photographs and the stage of DRComputers in Biology and Medicine 28 (1998) 225±2380010-4825/98/$19.00 # 1998 Elsevier Science Ltd. All rights reserved.PII: S001 0 -4825( 9 8 )00011 - 0PERGAMON* Corresponding author. Tel: +44-1224-681818 x52430; Fax: +44-1224-685645; E-mail: [email protected] according to a standard protocol [2, 3]. More accurate quanti®cation of the progress ofthe disease can be made by identifying and counting some of the lesions on the photographsuch as microaneurysms (MA). A positive correlation between the number of MAs and theprogression of the disease has been shown [4, 5].MAs, as seen in normal fundus photographs, are often not readily visualised and may beconfused with small dot haemorrhages. Fluorescein angiography increases the visibility of theselesions (a procedure involving the intravenous injection of a solution of sodium ¯uoresceindye). MAs are small saccular bulges in the walls of the retinal capillary vessels; they therefore®ll with dye during the angiography and are classically described as appearing as distincthyper¯uorescent round objects in the angiographic image. In practice, however, the appearanceof MAs can deviate from the classical description. In particular, some may appear inassociation with larger vessels, or they may be a part of a conglomeration of more than oneMA. Also, small MAs can look similar to other retinal pathologies or capillary crossings.Therefore the identi®cation of the MAs for counting is not trivial. A representative ¯uoresceinangiogram image of a diabetic patient is shown in Fig. 3a.Protocols have been developed for manually counting MAs [6], but they are time-consumingand are subject to observer error. They are, therefore, not used in current clinical practice.Automated computer techniques of digital image processing oer a fast, objective andrepeatable method of counting MAs, and this approach is currently being investigated inAberdeen.Our approach to the detection and quanti®cation of features of interest in an image is basedupon the classical model of computer vision. The process involves a number of fundamentalsteps: acquisition of the image, pre-rocessing, segmentation and classi®cation [7]. Quanti®cationis a trivial step at the end of the process.We have developed a fully automated image processing system that processes a single framefrom a sequence of angiographic photographs. The acquisition and pre-rocessing stages of thesystem are fully described by Spencer [8], and the segmentation has been updated byCree [9, 10]. The process is described brie¯y below.The output of the segmentation stage is a set of candidate objects that bear resemblance toMAs. A number of shape and grey-scale based features are measured on each candidate and itis these data features that are used to classify each candidate into one of two classes, namelyMAs and spurious objects. It is only the MA class that is of interest for quantifyingretinopathy.This paper describes our investigation into the classi®cation stage of the task. We haveapplied three classi®cation methods and compared their ecacy using receiver operatorcharacteristic (ROC) analysis [11]. The classi®ers tested were:1. An empirically-derived quantitative and logical rule-base (RBS).2. Linear discriminant analysis (LDA).3. A Learning vector quantiser (LVQ) arti®cial neural network (ANN).The ®rst method (RBS) is a manually derived classi®er and consequently requires extensiveeort. Furthermore, it can be dicult for a human observer to locate clustering relationshipsin high-dimensional data. The automated methods (LDA and LVQ) derive classi®ers almostA.J. Frame et al. / Computers in Biology and Medicine 28 (1998) 225±238226instantaneously with little or no manual intervention, and can handle multi-dimensional data.However, they are not so adept as a human observer at exploiting nonlinear or subtlerelationships in the data features. In clinical applications it is important to be able to developthe best classi®ers possible, both for accurate quanti®cation and also to avoid errors that mayhave clinical implications. Therefore the extra time expenditure in manually deriving an RBSmay be justi®ed if the RBS can outperform any automated classi®ers. In ophthalmic diagnosisall methods have been applied separately [12±14] but no studies have compared severalmethods on the same data.In Section 2, a brief description of the image processing strategy up to the point ofsegmentation of candidates for classi®cation is given. Section 3 outlines the methodology usedfor training and testing the classi®ers. We ®nd it useful to describe the methodology beforeintroducing the three classi®cation methods in
or
We will never post anything without your permission.
Don't have an account? Sign up