Towards clinical significance prediction using k^trans evidences in prostate cancer

摘要

Currently, K^trans coefficient maps have emerged to characterize tumor biology and treatment response. Salient localized coefficient on K^trans allows to detect and localize tumor regions from non-invasive MRI scanners. Nevertheless, such identified lesions on K^trans maps are highly variable and in much of the cases result in false positive indicators. In this work a set of labeled K^trans regions are processed into a supervised framework to correctly find true positive regions that are prostate cancer indicators. Three different algorithms were implemented to perform the classification: K-Nearest Neighbors (k-NN), Support vector machine (SVM), and Random forest (RaF). On a public dataset with 339 K^trans images on peripheral, transitional and anterior fibromuscular stroma regions, the SVM achieved an average accuracy of 80.83% with a ROC AUC of 0.68 on true evidence identification.