SALIENT: Frequency-Aware Paired Diffusion for Controllable Long-Tail CT Detection

Researchers introduce SALIENT, a frequency-aware diffusion model framework that improves detection of rare lesions in CT scans by generating synthetic training data in wavelet domain rather than pixel space. The approach addresses extreme class imbalance in medical imaging through controllable augmentation, achieving significant improvements in detection performance for low-prevalence conditions.

• →SALIENT performs diffusion in wavelet domain instead of pixel space, making it more computationally efficient for medical image synthesis.
• →The framework separates low-frequency brightness from high-frequency structural detail for better controllable generation.
• →Synthetic augmentation improved detection metrics with MS-SSIM increasing from 0.63 to 0.83 and FID decreasing from 118.4 to 46.5.
• →Optimal synthetic data ratios shift from 2x to 4x as labeled training data decreases, indicating seed-dependent augmentation needs.
• →The approach specifically targets long-tail detection scenarios where rare conditions suffer from precision collapse despite high AUROC.