From Correlation to Constraint: A Physicist’s Journey to Rethinking AI in Radiology
Atefeh Abdolmanafi, PhD
My training spans physics, computer science, and medical imaging, alongside a parallel practice in the visual arts. This interdisciplinary background shapes how I think about AI in imaging: as a system that should remain grounded in the physical processes that generate medical images.
When I first began working with deep learning in intravascular optical coherence tomography (IV-OCT), the results offered an exciting look into human arteries. Using just pattern recognition, deep learning models could segment various coronary lesions and plaques and classify them by type with high accuracy. Microstructural differences in intracoronary OCT images became separable in the learned feature space.
The performance was impressive. But I found myself asking a more fundamental question: Which patterns is the model learning? Is it learning true pathophysiology, or statistical regularities in pixel space, patterns in image intensities that may not correspond to clinically meaningful structures, such as imaging artifacts or acquisition-related variations that are unrelated to the underlying disease?
Medical images are not photographs. Rather, pixel intensities are defined by modality-specific physical interactions between imaging energy and biological tissue. These signals encode properties such as attenuation, scattering, interferometry, acoustic reflections, or magnetic relaxation, depending on the imaging modality. A deep learning model trained only on pixel intensities therefore will learn correlations between patterns of image intensities and the labels or outcomes provided during training. However, there is no guarantee that a model will learn the governing physical constraints or encode the physical process that generated the image and ultimately reveals the underlying pathology.
The distinction between pixel-level correlations and the underlying imaging physics became more important to me when I began focusing on problems where the output was not a diagnostic label (e.g., plaque type) but a measurement, such as abdominal aortic aneurysm progression or intracoronary plaque rupture risk prediction. These measurements carry direct clinical consequences. They must be reproducible, physically meaningful, and consistent across imaging systems and patient populations. I realized that the high segmentation accuracy that had originally inspired my interest in AI is not sufficient if the derived quantities lack stability and physiological validity.
Importantly, I studied physics first. I returned to these foundations when I began working on volumetric blood flow quantification using color flow ultrasound and was reminded how strongly imaging measurements are governed by physical laws. In this setting, the limitations of relying solely on statistical correlations became even clearer. Blood flow is not a static feature; it is dynamic motion encoded in phase evolution, signal decorrelation, and acoustic propagation. The backscattered signal reflects interference patterns governed by wave physics. In small vessels, partial volume effects introduce mixtures of stationary and moving scatterers within a single voxel. Coherence metrics shift, and phase relationships evolve.
Sometimes the values appear counterintuitive, but they remain consistent with physical principles, similar to how blue and yellow light produce white, while blue and yellow pigment produce green. The result may seem contradictory, but the physics is precise.
Correlation only identifies statistical associations, while constraint provides stability. Constraint defines what is physically allowable. The laws of physics do not change when datasets change. Models that ignore these constraints may perform well within familiar data distributions, but fail when acquisition parameters, hardware, or physiology vary.
This way of thinking was shaped not only by physics training but also by my artistic practice. In art, structure is not restrictive; it creates coherence. The act of composition determines how elements are arranged, proportion controls their size relationships, and balance ensures stability, organizing color, light, and spatial arrangement to guide the viewer’s perception. These constraints do not limit creativity; they allow meaning to emerge. Without structure, expression becomes noise. The same principle applies to imaging science. Without embedded physical constraints, deep learning models risk learning shortcuts tied to datasets rather than mechanisms tied to biology.
In my own coronary research, I no longer think of plaque rupture as a simple morphological event. Rather, it emerges from the interaction between geometry, material properties, and local flow dynamics. Structure and function are inseparable within a physically constrained system. If AI is to contribute meaningfully to such problems, it must move beyond correlation.
Radiology is fundamentally a measurement science governed by physics. AI systems that respect these constraints are more likely to generalize, remain stable, and produce clinically meaningful inferences.
Atefeh Abdolmanafi is a researcher with a master's degree in physics and a Ph.D. in computer science, specializing in medical image analysis, which she obtained from Université du Québec, Montreal, QC, Canada, in 2018. She has made advancements in cardiovascular imaging, with a focus on Intravascular Optical Coherence Tomography. Currently based at the University of Michigan's Department of Radiology in Ann Arbor, MI, USA, Dr. Abdolmanafi specializes in color flow ultrasound technologies, driving innovation in medical imaging. Committed to interdisciplinary excellence, she blends art and science to inspire creativity and enhance healthcare solutions.