Localized Adaptive Style Mixing for feature statistics manipulation in medical image translation with limited Data

Abstract

Medical image translation synthesizes missing modalities to aid clinical diagnoses, yet Generative Adversarial Networks (GANs) often overfit in limited data scenarios. This work introduces Localized Adaptive Style Mixing (LASM), a novel regularization strategy addressing this challenge. Unlike global statistical mixing, LASM segments discriminator feature maps into grids and blends localized high-order statistics (mean, variance, skewness, kurtosis) from reference and input images. This forces the discriminator to focus on structural content rather than style, effectively mitigating overfitting. Experiments on brain T1- to-CT, pelvic T1-to-CT, and T2-FLAIR synthesis tasks demonstrate that LASM-equipped GANs outperform state-of-the-art methods, achieving 54.84 FID (vs. 131.54 baseline) with only 10% training data. Notably, LASM requires no transfer learning and integrates seamlessly into existing frameworks. Our approach advances data-efficient medical image translation, particularly for rare diseases with scarce datasets. Code is available at here.