Multimodality tissue‐type imaging of prostate cancer

Abstract

Unlike many tumors, prostate cancer cannot be imaged reliably by any single, existing, conventional modality. This makes guiding biopsies, targeting therapy, and monitoring effects problematic. Therefore, our objective is to improve significantly the depiction of prostate cancers using multimodality, multifeature techniques combined with nonlinear classifiers such as artificial neural networks (ANNs) and support vector machines (SVMs). Our previous studies show encouraging potential for tissue typing and imaging based on ultrasound spectrum analysis using ANNs and SVMs that produce ROC‐curve areas of 0.84 to 0.88 compared to 0.64 for conventional ultrasound imaging. Similar studies by others have shown that magnetic‐resonance spectroscopy can produce comparable ROC‐curve areas based on the ratio of choline and creatine to citrate. Because these two modalities sense entirely distinct properties of tissue, i.e., microscopic mechanical properties by ultrasound and chemical properties by magnetic resonance, their combination should markedly improve classification performance in distinguishing between cancerous and noncancerous prostate tissue. We have undertaken preliminary studies incorporating these two modalities. These studies demonstrate the feasibility of coregistering features from each modality and generating new, combined‐modality images that exploit the attributes of both modalities.