Biomedical Signal Analysis: Contemporary Methods and Applications

12. Skin Lesion Classification

12 Skin Lesion Classification This chapter describes an application of biomedical image analysis: the detection of malignant and benign skin lesions by employing local information rather than global features. For this we will build a neural network model in order to classify these different skin lesions by means of ALA-induced fluorescence images. After various image preprocessing steps, eigenimages and independent base images are extracted using PCA and ICA. In order to use local information in the images rather than global features, we first add self-organizing maps (SOM) to cluster patches of the images and then extract local features by means of ICA (local ICA). These components are used to distinguish skin cancer from benign lesions. An average classification rate of 70% is achieved, which considerably exceeds the rate obtained by an experienced physician . These PCA- and ICA-based tumor classification ideas have been published in [21] and extend previous work presented in [19]. 12.1 Biomedical Image Analysis Many kinds of biomedical data, such as fMRI, EEG, and optical imaging data, form a challenge to any data-processing software due to their high dimensionality. Low-dimensional representations of these signals are key to solving many of the computational problems. Therefore, principal component analysis (PCA) commonly was used in the past to provide practically useful and compact representations. Furthermore, PCA was successfully applied to the classification of images [272]. One major de- ficiency of PCA is its global, orthogonal representation, which often cannot extract the intrinsic information of high-dimensional data. Independent component analysis (ICA) is a generalization of principal component analysis which decorrelates the higher-order moments of the input in addition to the second-order moments. In a task such as image recognition, much of the important information is contained in the higher-order statistics of the image. Hence ICA should be able to extract local feature like structures of objects, such as fluorescence images of skin lesions. Bartlett demonstrated that ICA outperformed the face recognition performance of PCA [18]. Finally, local ICA was 326 Chapter 12 (a) (b) (c) Figure 12.1 Typical fluorescence images of psoriasis (a), actinic keratosis (b), and a basal cell carcinoma (c). developed by Karhunen and Malaroiu to take advantage of the localized features in high-dimensional data [132]. Using Kohonen’s self-organizing maps [140], multivariate data are first split into clusters and then local features are extracted using ICA within these clusters. Here, we intend to classify skin lesions (basal cell carcinoma, actinic keratosis, and psoriasis plaques) through their fluorescence images (see figures 12.1 and 12.2). Even an experienced physician is unable to distinguish malignant from the benign lesions when fluorescence images are taken. For the Skin Lesion Classification 327 (a) (b) Figure 12.2 Nonfluorescence images of psoriasis (a), actinic keratosis (b), and basal cell carcinoma. sake of simplicity, we will just denote the diseases as malignant, since basal cell carcinoma is a skin cancer and actinic keratosis is considered a premalignant condition. 12.2 Classification Based on Eigenimages PCA is a well-known method for feature extraction and was successfully applied to face recognition tasks by Turk and Pentland [272], Bartlett 328 Chapter 12 et al. [17, 18] and others. Thereby images are decomposed into a set of orthogonal feature images called eigenimages, which can then be used for classification. A new image is first projected into the PCA subspace spanned by the eigenimages. Then image recognition is performed by comparing the position of the test image with the position of known images, using the reconstruction error as the recognition criterion. For a statistical analysis of the obtained results, hypothesis testing is used for a reliable classification. Calculation of the eigenimages Consider a set of m images x1, . . . , xm with each image vector xi = [xi(1), . . . , xi(N2 )] comprising N2 pixel values of the N × N image i. Merge the whole set of images into an N2 × m matrix X = [x1, . . . , xm] and assume the expectation value E {xi} of each image vector to be zero. Then the covariance matrix can be calculated according to Cov(X) = 1 m m  i=1 xix i = XX . A set of N2 orthogonal eigenimages ui can now be determined by solving the following eigenvalue problem: XX ui = Σui, (12.1) where Σ = diag [σ1, . . . , σN2 ] denotes the diagonal matrix with the variances σi of the projections ri...