AN AUTOMATIC CLASSIFICATION SYSTEM APPLIED IN MEDICAL IMAGES Bo QIU, Chang Sheng XU, Qi TIAN Institute for Infocomm (I2R), Singapore, 119613 {qiubo,xucs, tian}@i2r.a-star.edu.sgABSTRACTIn this paper, a multi-class classification system is developed for medical images. We have mainly explored ways to use different image features, and compared two classifiers: Principle Component Analysis (PCA) and Supporting Vector Machines (SVM) with RBF (radial basis functions) kernels. Experimental results showed that SVM with a combination of the middle-level blob feature and low-level features (down-scaled images and their texture maps) achieved the highest recognition accuracy. Using the 9000 given training images from ImageCLEF05, our proposed method has achieved a recognition rate of 88.9% in a simulation experiment. And according to the evaluation result from the ImageCLEF05 organizer, our method has achieved a recognition rate of 82% over its 1000 testing images. 1. INTRODUCTION With the fast development of modern medical devices, more and more medical images are generated, so that the demand becomes more and more urgent for automatically indexing, comparing, analyzing and annotating the huge volume of medical images. Medical images are a kind of medical evidence to patients and doctors. To interpret those medical evidences, generally doctors will use specialist vocabulary and natural language phrases, and relate them to some specific cases. It is difficult for some unskilled doctors but automatic annotation of medical images will do much help to them. For automatic annotation, which is a kind of automatic machine-based reasoning based on the evidence gathered, additional interpretive semantics must be attached to the image data. About this some methods have been explored in special domains, like the diagnosis of breast cancer [1]. But until now in a wider domain, there is no popular method for automatic annotation owing to the variety of medical images and the lack of relevant domain knowledge. So in this paper we simplify the problem into a multi-class classification problem, which means that the classification labels assigned to the classes are regarded as a simple annotation. According to [2], classification methods include parametric and nonparametric. With given training data, in this paper only parametric methods are considered, which includes Bayesian estimation (Maximum-Likelihood, Hidden Markov models, Expectation-Maximization, Fisher Linear Discriminant, Multiple Discriminant Analysis, etc.), Linear Discriminant functions (Perceptron Criterion Function, Relaxation Procedures, Minimum Squared-Error Procedures, PCA, SVM, Ho-Kashyap Procedures, etc.), Multi-layer Neural Networks, Stochastic methods (Simulated Annealing, Boltzmann learning, Evolutionary methods, etc.). The methods above have been applied successfully in many fields[2]. But until now the problem of medical images classification is a new and great challenge, because when compared with other classification problems, there are some particular difficulties in medical images: z Great unbalance between classes Figure 1 shows the size of each class in our database (see experiment part). It can be found that, class 6 has more than 500 samples, class 12 has more than 2,500 samples, class 34 has near 1,000 samples, while all the others are much less — the minimal class has only 9 samples. 20 largest classes occupy near 80% of the whole dataset. This unbalance makes many common classification methods unavailable. (a) Sizes vs. classes (b) Size percentages vs. classesFigure 1. Great unbalance between classes z Visual similarities between some classes (See Figure 2)Unlike the other image databases, for medical images, sometimes even skilled experts cannot find the differences between some classes visually. They may need to compare the images from different sources and refer to other medical examinations like blood. Figure 2. Visual similarities between some classes z Variety in one class and difficulty to define discriminative visual features (See Figure 3)Too many modalities vary in one class. To find a general visual feature for one class is often very difficult. In many cases, medical similarities are far away from visual similarities. 10451424403677/06/$20.00 ©2006 IEEE ICME 2006Figure 3. Variety in one class To face the difficulties mentioned above, based on our former work [7], PCA and SVM are chosen as classifiers in this paper. And different features from low-level to middle-level are considered. Our contributions are: z Construct a multi-class classification system for medical images; z Find the most efficient features for classification by designed simulation experiments (some training data are used to simulate testing data). 2. FEATURE SETS Feature extraction is a basic problem in image processing field. After reviewing 56 CBIR (content-based image retrieval) systems, in[3] a summary of low-level features are listed in 3 main categories: color, texture, and shape, plus a single features: layout. In[4][5] there are some similar categories of features. The feature ‘layout’ is the absolute or relative spatial position of the color. It may include low-resolution-pixel-map (LRPM), which is used in our method. LRPM is a down-scaled image of an initial one.In our system texture maps are calculated on both initial images and filtered images. Filtered images are generated from initial ones by filters like Gaussian, to minimize the influence of noises. Moreover, texture histogram is calculated on these texture maps. Figure 4 shows an example of textures and LRPM. image contrast anisotropy polarity LRPMFigure 4. An initial image and its feature maps Besides the low-level features like LRPM and texture, middle-level regional features such as Blob are also considered [6].Blob has been applied successfully in medical image retrieval in our past work [7]. Its parameters include: color, texture, area, length of long and short axes, rotation angle, Fourier decomposition parameters, etc. Because there are so many features available, feature selection becomes a key problem. The ‘best’ features should be the most distinguishing features, and invariant to irrelevant transformations of the input. Facing all kinds of features, it is extremely difficult to find out which are the best ones theoretically. The practical way is to select suitable features by simulation experiments,