Rad-pathomic deep learning models to assist radiologists in differentiating aggressive from indolent prostate cancer on MRI

点击翻译按钮获取中文摘要

PROJECT SUMMARY/ABSTRACT Prostate cancer is the second deadliest cancer for American men. MRI is increasingly used to guide prostate biopsies and has potential to spare 500,000 men/year from the side effects of invasive biopsies. Yet, subtle differences in MRI appearance of aggressive vs. indolent (non-lethal) cancer vs. benign tissue creates three problems: missed cancers, high rates of false positives, and only moderate inter-reader agreement among ra- diologists. Selective identification of aggressive and indolent cancers is imperative for reducing cancer death while minimizing side effects from unneeded biopsies. We propose to develop and use pathology-based (pathomic) MRI biomarkers in rad-pathomic deep learning methods to assist radiologists in detecting and localizing aggressive vs. indolent cancers on prostate MRI. In addition, our proposed method will be the first to localize aggressive and indolent cancers when they coexist (76% of index lesions). We performed four preliminary studies in our unique dataset of matched radiology and pathology images. First, we found a high agreement in labeling aggressive vs. indolent cancers between the automated method and two pathologists. Second, we developed pathomic MRI bi- omarkers from MRI features that correlate with features derived from pathology images. Third, we used the biomarkers in rad-pathomic deep learning models to detect cancer (AUC: 0.86) and aggressive cancer (AUC: 0.85) on MRI. Fourth, we showed that combining radiologists and the rad-pathomic deep learning models helped identify 14% more aggressive cancers missed by radiologists. Three innovations will improve the localization of aggressive vs. indolent cancers on prostate MRI. First, we will develop 3D RAPSODI, a novel 3D registration method for 3D reconstructed MRI and pathology images to eliminate the need for slice-to-slice correspondences and map cancer labels from pathology onto MRI. Second, we will leverage our correlation learning method to identify pathomic MRI biomarkers. Third, we will use deep learning models to assist radiologists in localizing aggressive cancer on MRI. Our multidisciplinary team is uniquely positioned to test whether: (Aim 1) pathomic MRI biomarkers empha- size the visual differences of aggressive vs. indolent cancers on MRI; (Aim 2) rad-pathomic deep learn- ing models can reliably and automatically distinguish aggressive from indolent prostate cancers on MRI, and (Aim 3) radiologists assisted by deep learning models have increased detection accuracy and inter-reader agreement than unassisted radiologists. Impact: Our proposed rad-pathomic deep learning models have the potential to improve prostate cancer care in three ways: 1) detecting and targeting aggressive cancers that are currently missed in ~50,000 men/year; 2) eliminating up to 500,000 unnecessary biopsies/year in men with no cancer or indolent cancers; and 3) reduc- ing the number of biopsy samples needed to detect aggressive cancers (1-2 vs. 12-18 currently).