Brain tumor classification on intraoperative contrast-enhanced ultrasound

Contrast-enhanced ultrasound (CEUS) imaging of tissue perfusion is based on microbubble echo detection. CEUS can visualize tumors based on local perfusion variations. The acquired video data are qualitatively interpreted by subjective visualization in clinical practice. An automated CEUS classifier was developed for intraoperative identification of tumor tissue and especially tumor borders. Support vector machines (SVM) were trained using CEU data sets to differentiate tumor and non-tumor tissue in glioblastoma patients. The classification was based on features derived from model functions approximated to time courses for each pixel in the video data. Classification performance was evaluated with single and cross- patient training data sets. The minimum mean classification error (14.6 %) with single patient data set training was achieved by SVM training using a sigmoid combination of model function parameter sets. A comparison of different model functions showed that the minimum average classification error (17.4 %) in a cross-validation study with 13 patients was achieved with the sigmoid model using an automatic relevance detection kernel. CEUS-based classification map images derived from approximated model functions can be generated with moderate accuracy and have significant potential to support intraoperative decisions concerning glioblastoma tumor borders.