A New paradigm for Ultrasound-Based Tissue Typing in Prostate Cancer
MetadataShow full item record
Prostate cancer is the most common malignancy among men. The gold standard clinical diagnosis method for prostate cancer is histopathologic analysis of biopsy samples acquired under ultrasound guidance. However, most prostate tumors lack visually distinct appearances on medical images. Therefore, pathologically significant cases of cancer can be missed during biopsy, resulting in false negative or repeated trials. The goal of our research is to augment ultrasound-guided prostate biopsy by adding tissue typing information that can be used for targeted biopsies. As a new paradigm in tissue typing, we hypothesize and demonstrate that if a specific location in tissue undergoes sequential interactions with ultrasound, the time series of echoes, which we call radiofrequency (RF) time series, would carry ``tissue typing'' information. We provide a potential physical explanation for this phenomenon and justify it based on computer simulations of the ultrasound probe and scattering media. We also report laboratory and animal studies that illustrate the utility of the method. We rely on a set of seven spectral and fractal features extracted from RF time series for tissue typing. To show the clinical value of the proposed approach, we report an ex-vivo study involving 35 patients in which the utility of RF time series features for detection of prostate tumors is confirmed. The outcomes are validated using histopathologic disease distribution maps provided for the studied specimen. We show that the RF time series features are powerful tissue typing parameters that result in an area under receiver operating characteristic (ROC) curve of 0.87 in 10-fold cross validation for diagnosis of prostate cancer. They are significantly more accurate and sensitive than spectral features extracted from single RF frames, and also B-scan texture features (area under ROC curve of 0.78 and 0.72, respectively). A combination of these three categories of features results in a feature vector that provides an area under ROC curve of 0.95 in 10-fold cross-validation and 0.82 in leave-one-patient-out cross validation for diagnosis of prostate cancer. Using this hybrid feature vector and support vector machines, we create cancer distribution probability maps that highlight areas of tissue with high risk of cancer.