PhaseAT improves medical image AI across scanner differences
PhaseAT: Fourier Phase Adversarial Training for Medical Image Domain Generalization
Computer Vision and Pattern RecognitionMachine Learning
Summary
Medical image AI programs often struggle when images come from different machines or hospitals than they were trained on. The authors developed a new training approach called PhaseAT that changes key image structures to make the AI focus on important shapes rather than textures. This method helps the AI work better when faced with new or varied medical images, improving reliability. They tested PhaseAT on two medical image datasets and saw significant improvement compared to other methods.
What this means in practice
- •For medical device developers: Build AI models for medical imaging that maintain accuracy across different scanning machines and hospitals.$Commercial implications: Enables reliable AI-enabled medical imaging products usable in diverse clinical settings with varying equipment.
- •For automated quality control teams: Improve image analysis tools to handle variations caused by production line cameras or sensors in pharmaceutical manufacturing.
Authors
Ahmed Sharshar, Asif Hanif, Naveen Kumar Kummari, Mohammad Yaqub, Mohsen Guizan
Abstract
Reliable clinical deployment of deep medical image models is hindered by distribution shifts across scanners, sites, and acquisition protocols. Existing domain generalization (DG) methods often focus on style or intensity diversification, but they can still leave networks dependent on domain-specific texture correlations. Inspired by evidence that Fourier phase encodes semantic structure, we introduce PhaseAT, a phase-aware adversarial training framework for medical DG. PhaseAT forms phase-perturbed training views in the Fourier domain by iteratively updating a bounded phase perturbation while keeping the amplitude spectrum unchanged, thereby stressing spatial organization under matched appearance statistics. Perturbations are applied only to the luminance channel in YCbCr color space to avoid chromatic artifacts. Additionally, a simple phase-saliency mask concentrates updates on the most influential frequencies. The model is trained with a weighted combination of losses on clean and phase-perturbed samples, supporting both single-source and multi-source DG. We validate our method on two challenging medical datasets and demonstrate that PhaseAT achieves over 20% improvement in single-source domain generalization, outperforming several state-of-the-art DG methods. The code implementation is available at: https://github.com/ahmed-sharshar/PhaseAT.