Abstract
Magnetic Resonance Imaging (MRI) is among the most diagnostically powerful tools in modern medicine, yet it remains constrained by significant limitations in speed, cost, and accessibility. A standard MRI exam can take between 20 and 60 minutes, contributing to patient discomfort, motion-related imaging artifacts, high operational costs, and limited patient throughput in clinical settings. These limitations are not merely technical inconveniences; they carry real consequences for healthcare accessibility, with a substantial portion of the global population lacking reliable access to MRI technology. Artificial intelligence (AI) has emerged as a compelling avenue for addressing these challenges, particularly through its application to the image reconstruction pipeline. AI-based reconstruction methods, especially those leveraging deep learning, have demonstrated the ability to produce high-quality diagnostic images from sparser data than conventional methods require, enabling shorter scan times and reduced hardware costs. However, the technical capacity of these tools does not guarantee their clinical adoption. In healthcare, where the stakes of failure are exceptionally high, a technology must also earn the trust of patients, radiologists, and clinical institutions before it can become a stable part of clinical practice. Together, improving the technical performance of AI MRI reconstruction and understanding the social conditions that shape its clinical adoption define the central problem addressed by this thesis portfolio.
The technical component of this portfolio investigates a novel approach to AI-based MRI image reconstruction using a trained diffusion model as a learned regularizer within the reconstruction pipeline. Conventional MRI reconstruction methods are based on well-established physics models that, while reliable, require dense data sampling and therefore long scan times. Recent deep learning approaches have sought to address this by training neural networks to reconstruct images from undersampled data, but these methods often struggle to balance image fidelity with noise suppression and generalization across different scan settings. This project proposes to constrain the reconstruction optimization problem using a pre-trained diffusion model, which acts as a powerful prior capturing the learned distribution of high-quality MRI images. By incorporating this diffusion-based regularizer, the reconstruction pipeline can enforce realistic image structure during optimization, producing sharper and more diagnostically accurate images from fewer acquired data points than current methods allow. The approach is evaluated against standard reconstruction baselines on image quality metrics and diagnostic feature preservation, with the aim of demonstrating that diffusion model regularization provides a meaningful and measurable improvement in reconstruction performance. If successful, this method has the potential to reduce scan times substantially while maintaining or improving image quality, directly addressing one of the core technical barriers to broader MRI accessibility.
The STS research component of this portfolio examines the techno-social factors shaping the emergence of AI MRI reconstruction as a stable clinical practice, drawing on the Social Construction of Technology (SCOT) framework. SCOT posits that technologies are stabilized not simply through technical performance, but through a process of social negotiation among relevant groups whose interpretations and concerns shape how a technology is understood, adopted, or rejected. This paper identifies three relevant social groups: patients, radiologists, and clinical institutions, and analyzes the distinct problem-spaces each group brings to AI MRI reconstruction. For patients, the primary concern is the lengthy and uncomfortable nature of conventional MRI scanning; AI reconstruction addresses this directly by enabling faster scan times and greater tolerance for patient movement, and clinical study data shows patients rate AI-assisted MRI exams significantly more favorably for duration than conventional exams. For radiologists, the key concern is whether AI reconstruction matches or improves upon the diagnostic image quality of conventional methods, a standard AI has been shown to meet while also offering meaningful denoising advantages. For clinical institutions, the economic case for AI MRI is compelling, as it reduces hardware costs and increases patient throughput, but regulatory gaps persist, with most FDA-cleared AI radiology devices receiving streamlined approval with limited prospective clinical testing. The paper further examines the black-box problem, arguing that while AI opacity poses a genuine barrier to patient trust, the concept of infrastructural invisibility, the tendency of trusted clinical technologies to become socially transparent without losing legitimacy, offers a viable pathway to closure for AI MRI reconstruction.
Together, these two projects address the same underlying challenge from complementary directions. The technical project works to close the performance gap between AI reconstruction and conventional MRI, establishing the kind of relative advantage that is a prerequisite for meaningful technology adoption. The STS project maps the social terrain the technology must navigate in order to be accepted as a clinical standard. A technically superior reconstruction method will still fail to achieve clinical closure if it cannot satisfy the trust requirements of radiologists, the comfort needs of patients, and the regulatory expectations of institutions. Conversely, understanding the social conditions for closure is most valuable when the technology in question genuinely delivers on its technical promises. This portfolio argues that AI MRI reconstruction is approaching the conditions necessary for stable clinical adoption, but that persistent regulatory gaps and unresolved liability questions remain the most significant obstacles standing between a promising technology and widespread clinical legitimacy.
