Prostate cancer (PCa) is the most common form of cancer in men in the Western countries. A non-invasive method localizing PCa may reduce the number of biopsies by targeting the biopsy needle to a suspect tumor. Moreover, it could assist in efficient application of focal treatments. Contrastultrasound dispersion imaging (CUDI) has recently been presented as a non-invasive method for early detection and localization of aggressive forms of prostate cancer. A classifier trained by combining features based on contrast ultrasound could further improve the accuracy of localizing the tumor. In this work, support vector machine (SVM) classification algorithm was implemented with not only the dispersion-based features obtained from CUDI, but also other perfusion-based features estimated from dynamic contrast-enhanced ultrasound. Training and testing of the classifier was based on histopathological results obtained from radical prostatectomy (excision of the prostate) in 6 patients. To determine the optimal combination of features, all the possible combinations were compared. Meanwhile, a 10- fold cross-validation by comparison with histology results was performed to optimize the SVM parameters on a pixel basis. Nine feature combination shows the best classification performance among all the combinations. Leave-one-out cross validation on an imaging plane basis was also used to evaluate the classification performance in practice. The SVM classification with multiple features improves the accuracy by comparison with the threshold classification of the individual features, encouraging further validation with a larger patient group.
Classification and validation of contrast ultrasound dispersion maps compared with histology results
Wu, K. (Auteur). 31 okt. 2014
Scriptie/Masterproef: Master