2G-6 Classification of Parotid Gland Tumors using Sonohistology

In this paper results from a clinical study on differentiating between various types of parotid gland tumors using computerized tissue characterization (Sonohistology) are presented. Complex baseband ultrasound data have been acquired during the common examinations of patients who were scheduled to have parotid surgery shortly after the acquisition. Data of 123 benign and malignant parotid-gland alterations have been included in the study. For data acquisition, a conventional diagnostic ultrasound scanner controlled by custom software running on a laptop computer was used. Tumors were manually contoured in the B-Mode images. Acquired data were stored on an external PC and subdivided into numerous regions of interest (ROI). For each ROI, a set of tissue characterizing spectral and texture features was calculated. Moreover, Fourier descriptors have been calculated from the contours of the lesions to characterize differences in shape of certain kinds of tumors. Training data have been generated from the manually contoured areas. For classification, the training data have been divided in up to four subclasses. The final classification was done using two target classes. The first class included all cases for which a surgical treatment was definitely necessary. The second class included all cases that did not necessitate a surgical treatment. The best feature set was processed by a classification system. For classification, the maximum likelihood measure was used. Classification was done by total cross validation over cases. The best feature set was found by sequential forward selection and included two spectral features (attenuation and slope), two first order texture feature (squared signal to noise ratio and kurtosis), two measures from the cooccurrence matrix (sum variance and variance of sum of squares) and two Fourier descriptors. The receiver operating characteristics curve area was A_ROC = 0.86