Introduction
Deep learning has transformed medical image analysis (MIA), yet it faces a significant "black box" problem. Traditional models often require massive amounts of labeled data, struggle with interpretability, and frequently ignore the physical laws—such as fluid dynamics or electromagnetism—that govern how medical images are formed. This survey explores Physics-Informed Medical Image Analysis (PIMIA), a burgeoning field that integrates these laws directly into neural network architectures.
The authors argue for a paradigm shift: rather than forcing models to learn everything from scratch, we should provide them with the "cheat sheet" of physical principles we already know. This approach addresses data scarcity and prevents "hallucinations" that could be fatal in clinical settings.
Methodology: The PINN Framework
The core technique in PIMIA involves modifying the standard neural network loss function. In addition to measuring error against ground-truth labels ($L_{data}$), a second term ($L_{eqns}$ or $L_{PDE}$) is added to penalize outputs that violate specific physical laws, such as the Navier-Stokes or Helmholtz equations.
The survey categorizes these integrations into three types of "biases":
- Observational Bias: Incorporating physics-simulated data or raw sensor measurements (like MRI k-space) as input.
- Learning Bias: Using the loss function to enforce physical constraints, such as ensuring mass is conserved in blood flow.
- Inductive Bias: Designing network architectures that inherently mirror physical processes (e.g., Hamiltonian Neural Networks).
Key Insights and Results
A meta-analysis of over 80 works reveals that physics integration provides varying levels of efficacy depending on the task. The authors introduced a novel "Improvement Score" ($S$) to quantify these gains relative to non-physics baselines:
- Super-Resolution (44.27% improvement): Physics-based kernels help fill in high-frequency details more accurately than simple interpolation.
- Predictive Modeling (29.67% improvement): Particularly effective in hemodynamics, where Navier-Stokes equations prevent unphysical flow predictions.
- Registration (5.5% improvement): Biomechanical models ensure realistic tissue deformation during image warping.
Notably, high-level tasks like Segmentation saw only marginal gains (2.85%), suggesting that these are currently treated primarily as pattern recognition tasks rather than physical ones.
Limitations and Future Outlook
Despite its promise, PIMIA faces several "open frontiers." Overregularization is a concern, as rigid physical constraints might cause models to ignore real clinical anomalies that do not perfectly fit simplified mathematical models. There is also a significant need for standardized benchmarking (a "PIMIA-Bench") and models that generalize across different hospital hardware. Finally, the authors suggest that physics is most powerful when applied to raw sensor data (k-space in MRI or sinograms in CT), which remains a challenge due to data accessibility hurdles.
Summary updated