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

A research team led by Yifan Li has introduced SALIENT, a novel AI framework designed to solve a critical bottleneck in medical imaging: detecting extremely rare lesions in whole-body CT scans. The problem, known as 'long-tail detection,' is plagued by extreme class imbalance where positive examples are vastly outnumbered by normal tissue, causing AI models to suffer precision collapse despite high overall accuracy. SALIENT addresses this by generating high-quality, controllable synthetic CT data for augmentation, but with a key architectural innovation—it performs diffusion in the wavelet domain rather than the traditional pixel space. This allows the model to explicitly separate and control low-frequency attributes like brightness from high-frequency structural details, leading to more realistic and attribute-controllable synthetic volumes.

The technical core of SALIENT involves a mask-conditioned wavelet-domain diffusion model paired with a 3D VAE for generating diverse lesion masks and a semi-supervised teacher for producing pseudo-labels. This frequency-aware approach enables disentangled optimization for target and background attributes, resulting in significant gains in generative realism. In downstream evaluation, training detection models with SALIENT-augmented data yielded substantial improvements in Area Under the Precision-Recall Curve (AUPRC), especially for the rarest conditions. The research also identified a seed-dependent augmentation regime, finding the optimal synthetic data ratio shifts from 2x to 4x as the amount of initial real labeled data decreases. This work demonstrates that computationally efficient, frequency-aware synthesis can provide a precision rescue for AI diagnostics in low-prevalence medical scenarios where collecting real data is impractical.