SVM-Based Characterization of Focal Kidney Lesions from B-Mode Ultrasound Images.

Abstract

Characterization of benign and malignant focal kidney lesions such as angiomyolipomas (AMLs) and renal cell carcinomas (RCCs) from ultrasound images is an intimidating task for radiologists due to their enormously overlaying sonographic presences. In the present work the main focus is on the aspect of variations in texture patterns shown by focal kidney lesions. In order to visualise these textural disparities, texture features are calculated using different methods namely statistical features and spectral texture features. These texture features have been calculated from region of interest (ROIs) extracted from each image within the lesion. The study is performed on 23 ultrasound kidney images with 14 AML lesions and 9 RCC lesions. For classification task support vector machine classifier has been used to classify data based on analysis. It has been seen that by combining Gray level run length statistics and Fourier power spectrum features produces the maximum accuracy of 84% with ICA values of 100 % and 60 % for RCC and AML lesions respectively.