6C-2 Differential Diagnosis of Parotid Gland Lesions Using Spatially Fused Sonohistologic Features

In an ongoing clinical study a sonohistology system is developed and evaluated towards its ability to perform computerized differential diagnosis of parotid gland lesions. First order statistics are used to calculate fused features from spatially resolved parameter images. Thereby, characteristics of patterns representing the type of lesion are quantified. 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 benign and malignant parotid-gland alterations originating from 135 patients 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. Lesions were manually contoured in the B-mode images. Acquired data were stored on an external PC. Fused features were calculated offline. From a large number of fused features, a best performing subset is chosen by a selection algorithm to form a feature vector representing each case. The best feature set was used to classify each case using leave-one-out cross validation. Two different classifiers have been used for comparative reasons: a probabilistic neural network based on radial basis functions, and a maximum likelihood classifier, yielding areas under the ROC- curve of 0.85 and 0.91 with standard errors of 0.04 and 0.03, respectively. The system can be adjusted to reach a sensitivity of 1 to catch all positive cases, leaving a remaining maximal specificity of 0.55. Therefore, the system can be used to optimize treatments of parotid gland lesions and to reduce the number of unnecessary surgical interventions.