Standardizing Medical Images at Scale for AI

A team led by Bahram Jalali at UCLA, with collaborators from the University of Tokyo, has published a breakthrough paper on arXiv introducing PhyCV, a physics-based framework designed to solve a critical bottleneck in medical AI: data heterogeneity. Deep learning models for analyzing X-rays, MRIs, and histopathology slides often fail when deployed at new hospitals because of differences in imaging hardware, staining protocols, and lighting. PhyCV addresses this by modeling medical images not as simple pixels, but as spatially varying optical fields. It applies a deterministic transformation inspired by physical optics—virtual diffractive propagation followed by coherent phase detection—to strip away irrelevant, institution-specific variations in color and illumination while preserving the diagnostically crucial textures and structures.

The impact is substantial and quantifiable. When tested on the challenging Camelyon17-WILDS benchmark for metastatic breast cancer detection in lymph node tissue, standard AI models trained with Empirical Risk Minimization achieved only 70.8% accuracy on out-of-distribution data from unseen hospitals. Preprocessing the same images with PhyCV before training boosted that accuracy to 90.9%, a 20-percentage-point improvement that matches or exceeds more complex domain-generalization techniques. Crucially, the PhyCV transform is parameterizable, differentiable, and adds negligible computational cost. It can be used as a fixed preprocessing 'data refinery' or integrated directly into an end-to-end neural network pipeline. This work, grounded in first-principles physics, provides a powerful tool to harmonize disparate medical datasets, paving the way for more robust, interpretable, and reproducible AI systems that can reliably work across global healthcare networks.