Physics‑based PhyCV preprocessing reduces stain and scanner differences and raises out‑of‑hospital breast cancer accuracy from 70.8% to 90.9%
This paper introduces PhyCV, a physics‑based preprocessing method that aims to make medical images more consistent before they are fed into AI. The core idea is to treat an image like an optical field and run it through a virtual optical propagation step followed by coherent phase detection. That process removes non‑semantic differences such as color, stain, and lighting while keeping the texture and structures doctors care about.
The authors describe a family of algorithms that implement this idea. They transform a real image into a complex‑valued field in the frequency domain, apply a small spectral phase change that mimics diffraction or dispersion, and then detect the resulting spatial phase. The output is a phase image made of “phixels” that highlights edges and fine textures. Mathematically, under a small‑phase approximation, the operation behaves like a Laplacian‑style filter that equalizes contrast and illumination while enhancing structural features.
To test the approach, the team applied PhyCV to histopathology images from the Camelyon17‑WILDS benchmark, a dataset built to test how well models generalize across hospitals. With PhyCV preprocessing, out‑of‑distribution breast cancer classification accuracy rose from 70.8% for a standard Empirical Risk Minimization baseline to 90.9%. The authors report that this matches or exceeds the gains from larger data‑augmentation or domain‑generalization approaches, and that PhyCV adds negligible computational cost.
PhyCV is presented as a practical data‑refinery for medical imaging. Because the transform is rooted in optical physics, it is physically interpretable, parameterizable, and differentiable. That means it can be used as a fixed preprocessing step before training and inference, or it can be inserted into an end‑to‑end learning pipeline where gradients flow through it during training. The paper also points to other PhyCV uses such as super‑resolution of magnetic resonance (MRI) images and vessel extraction in retinal images, which illustrate the method’s versatility.