A Path Towards Leveraging AI to Enable Point-of-Care MR Devices for Disease Detection

New York University School Of Medicine, New York NY

Investigators

Abstract

Project Summary Prostate Cancer (PCA) is the second most prevalent malignancy and a leading cause of mortality among men in the United States. The urgency for early identification of Clinically Significant Prostate Cancer (CSPCA) within the at-risk population cannot be overstated, as it holds the key to significantly enhancing clinical outcomes. Yet, the current standard-of-care for CSPCA surveillance, the Prostate Specific Antigen (PSA) test, is plagued by low specificity, resulting in a cascade of unnecessary advanced diagnostic imaging and invasive biopsy procedures, inflicting a substantial wasteful cost burden on the health system and subjecting patients to avoidable trauma. Despite the established efficacy of Magnetic Resonance Imaging (MRI) in diagnosing CSPCA, it is not used as a first-in-line tool to identify patients with CSPCA at the population level. The reluctance to embrace MRI is rooted in the requirement to generate high-fidelity images, imposing inherent challenges – the need for expensive scanners installed in specialized imaging centers, executing complex and slow protocols tailored to acquire copious high- quality k-space data, and the dependence on sub-specialized radiologists for image interpretation. Driven by our mission to democratize MR diagnostics for widespread surveillance of critical illnesses like CSPCA on a population scale, we challenge the conventional wisdom that high-fidelity images are necessary for accurate disease inference. Instead, we advocate for inferring disease presence directly from a meticulously curated subset of degraded k-space data. We hypothesize that modern Machine Learning (ML) models can accurately infer the presence of the disease (a binary decision - which is all that may be required for disease surveillance) using only a minimal amount of carefully selected degraded k-space data. This carefully chosen degraded data, that would otherwise be insufficient to generate a diagnostic-quality image, can be swiftly acquired using inexpensive and accessible scanning devices, thereby paving the way for making MR technology accessible for disease identification at the Point-of-Care (POC). In this proposal we will first establish the feasibility of ML models to accurately detect CSPCA directly from a fixed under-sampled k-space data, without images (Aim 1). Subsequently, we will determine the minimum quantity of k-space data required for accurate disease detection, irrespective of image quality, by proposing an innovative end-to-end ML methodology that identifies a diminutive subset of k-space containing ample information for precise CSPCA inference and uses it to draw conclusions about CSPCA presence (Aim 2). Finally, we will aim to show that CSPCA can be accurately inferred from the degraded k-space data acquired using low-field scanners, without images (Aim 3). Findings of this project will establish the groundwork for affordable and accessible MR devices that maintain accuracy in detecting CSPCA. These devices, capable of operating closer to the POC for disease surveillance, will promote equitable access to MR diagnostics beyond specialized environments, revolutionizing disease surveillance on a broader scale.