Tuning kernel function parameters of support vector machines for segmentation of lung disease patterns in high-resolution computed tomography images

High-resolution computed tomography (HRCT) produces lung images with a high level of detail which makes it suitable for diaunosis of diffuse lung diseases. Segmentation of abnormal lung patterns is a necessary stacre in the construction of a computer-aided diagnosis system. We interpret lung patterns as textures and apply a texture classification technique for segmentation of lung patterns. The wavelet transform is used to extract texture features and then the Support Vector Machines (SVM) machine learning algorithm is applied to texture classification. The parameters of the SVM play a crucial role in the performance of the algorithm. We apply gradient-based optimization of the radius/margin bound of a generalization error to choose parameters of the SVM algorithm. This approach is more efficient in terms of the required number of SVM training cycles than the commonly used method of finding the optimal parameters which is based on sampling the parameter space and choosing the parameter combination which produces the lowest test error. We assess the applicability of optimization of the radius/margin bound to tuning SVM parameters for the problem of segmentation of lung pattern textures in HRCT images. Results of experiments indicate that this method chooses parameters which are comparable to the parameters obtained using test error in terms of classification accuracy, employing fewer training cycles.