Physics-Informed Neural Networks Improve Medical Image Registration with Adaptive Mechanical Constraints

When doctors need to track how organs move and deform over time—comparing a scan taken today against one from months ago—they rely on image registration, a computational process that aligns two images and maps the deformations between them. It's essential work: tracking tumor growth, monitoring organ function, planning surgery. But the standard approach has a blind spot. Most registration systems apply the same mechanical rules everywhere in the image, treating bone the same way they treat soft tissue, assuming uniform behavior across the entire anatomical landscape. In regions where the image is faint or poorly defined, this one-size-fits-all constraint becomes the only guide the algorithm has, and it often gets the deformation wrong.

Researchers have now embedded the actual physics of human tissue directly into the neural networks that perform this registration. Instead of imposing uniform constraints, the new framework teaches the network to recognize what kind of tissue it's looking at and apply the appropriate mechanical rules locally. Bone stays rigid. Soft tissues deform elastically. At the boundaries between different structures—where organs slide past each other—the network learns to capture that shearing motion. The key innovation is that these physical laws aren't bolted on as a post-processing step; they're woven into the training process itself, so the network internalizes them and applies them automatically when it encounters new patient data it has never seen before.

The team validated the approach on three-dimensional abdominal images, both synthetic test cases and real scans from actual patients. The results showed that a network trained with these mechanics-informed constraints successfully transferred its learned behavior to new patients within the same cohort. Hard tissues remained appropriately rigid. Tissue interfaces captured the sliding motions that actually occur in the body. The deformations the network produced were more consistent with how those structures actually move and bend in living anatomy. There was a trade-off: the intensity-based alignment—the pixel-level matching between images—was slightly less perfect than it would have been with unconstrained registration. But the cost was small and quantifiable, a deliberate choice to gain mechanical realism.

What makes this work is the principle of physics-informed machine learning: rather than asking a neural network to learn patterns purely from data, you give it the physical laws as part of its objective function. The network still learns from examples, but it's learning within a framework that respects the actual behavior of tissue. At inference time, when the network encounters a new patient's scan, it needs no additional information about that patient's anatomy or mechanical properties. The constraints it learned during training generalize automatically, because they're based on universal principles of solid mechanics, not on memorized patterns from the training set.

The practical implication is significant. Medical imaging often involves poorly contrasted regions where the image signal is weak and regularization constraints become the dominant source of information about how structures deform. A constraint that ignores tissue type will produce implausible deformations in these regions. A constraint that respects biomechanics will produce deformations that align with how the body actually works. For clinicians comparing scans over time, or for surgical planning, or for tracking disease progression, this difference between plausible and implausible deformation can matter. The code has been made publicly available, opening the approach to other researchers and clinical teams working with three-dimensional medical images.