Prostate cancer (PCa) is the most diagnosed form of non-cutaneous cancer in US men. The selection of patients who require immediate treatment from those suitable for active surveillance currently relies on non- specific and inaccurate measurements. A method that allows clinicians to more confidently discriminate clinically relevant from non-life-threatening tumors is needed to improve patient management. Multiparametric magnetic resonance imaging (mpMRI) is the preferred non-invasive imaging modality for characterizing primary PCa. However, its accuracy for detecting clinically significant PCa is variable. We propose to address this limitation by combining mpMRI with positron emission tomography (PET) with a PCa-specific radiotracer and using advanced multimodal machine learning models (i.e. radiomics and deep learning) to characterize tumor aggressiveness based on the imaging data. Recently, scanners capable of simultaneous PET and MR data acquisition in human subjects have become commercially available. An integrated MR-PET scanner is the ideal tool for comparing MR and PET derived image features to identify those that provide complementary information and build a hybrid PET-mpMRI model that most accurately identifies clinically significant tumors. While this novel technology allows the acquisition of perfectly coregistered complementary anatomical, functional and metabolic data in a single imaging session, a new challenge needs to first be addressed to obtain quantitatively accurate PET data. In an integrated MR-PET scanner, the information needed for PET attenuation correction (AC) has to be derived from the MR data and the methods currently available for this task are inadequate for advanced quantitative studies. We have formed an academic-industrial partnership to accelerate the translation of multimodal MR-PET machine learning approaches into PCa research and clinical applications by addressing the AC challenge and validating machine learning models for detecting clinically significant disease against gold standard histopathology in patients undergoing radical prostatectomy. Specifically, we will: (1) Develop and validate an MR-based approach for obtaining quantitatively accurate PET data. We hypothesize that attenuation maps as accurate as those obtained using a 511 keV transmission source ? the true gold standard for PET AC ? will be obtained; (2) Identify the multimodal radiomics model that most accurately predicts PCa aggressiveness. We hypothesize that the diagnostic accuracy of this approach will be superior to that offered by the stand-alone modalities; (3) Evaluate radiomics and deep learning approaches for predicting pPCa aggressiveness. We hypothesize that machine learning approaches will achieve a higher predictive accuracy when applied to data acquired simultaneously than sequentially.
A better method to non-invasively characterize primary prostate cancer is needed to improve patient management. Extracting additional information from multimodality quantitative MR-PET data using machine learning approaches is expected to result in better diagnostic performance. In this work, we propose to accelerate the translation of quantitative MR-PET to prostate cancer research and clinical applications. In particular, we will develop and validate an MR-based attenuation correction approach to guarantee that quantitatively accurate PET data are obtained in an integrated MR-PET scanner and then use machine learning approaches to characterize the aggressiveness of the tumors in patients undergoing radical prostatectomy.
|Torrado-Carvajal, Angel; Vera-Olmos, Javier; Izquierdo-Garcia, David et al. (2018) Dixon-VIBE Deep Learning (DIVIDE) Pseudo-CT Synthesis for Pelvis PET/MR Attenuation Correction. J Nucl Med :|