Title:A Multiresolution Clinical Decision Support System Based on Fractal Model Design for Classification of Histological Brain Tumours

Abstract:Tissue texture is known to exhibit a heterogeneous or non-stationary nature, therefore using a single resolution approach for optimum classification might not suffice. A clinical decision support system that exploits the subband textural fractal characteristics for best bases selection of meningioma brain histopathological image classification is proposed. Each subband is analysed using its fractal dimension instead of energy, which has the advantage of being less sensitive to image intensity and abrupt changes in tissue texture. The most significant subband that best identifies texture discontinuities will be chosen for further decomposition, and its fractal characteristics would represent the optimal feature vector for classification. The performance was tested using the support vector machine (SVM), Bayesian and k-nearest neighbour (kNN) classifiers and a leave-one-patient-out method was employed for validation. Our method outperformed the classical energy based selection approaches, achieving for SVM, Bayesian and kNN classifiers an overall classification accuracy of 94.12%, 92.50% and 79.70%, as compared to 86.31%, 83.19% and 51.63% for the co-occurrence matrix, and 76.01%, 73.50% and 50.69% for the energy texture signatures, respectively. These results indicate the potential usefulness as a decision support system that could complement radiologists diagnostic capability to discriminate higher order statistical textural information, for which it would be otherwise difficult via ordinary human vision.