Medical Super-Resolution Reconstruction Network for Downstream Tasks Based on Wavelet Analysis

A single low-resolution medical image may yield multiple feasible high-resolution reconstructions. Current mainstream super-resolution (SR) methods mainly optimize objective metrics like PSNR, while recent perceptual quality-enhancing algorithms still overlook reconstructed images impact on downstream diagnostic tasks. PSNR-optimized images are overly smooth and lack diagnostic details, while perception-prioritized ones may introduce “hallucination” textures, both interfering with automated diagnosis or causing misdiagnosis. To address this, we propose an SR framework tailored for downstream diagnostics. Without integrating downstream models, it achieves task-adaptive reconstruction via a constrained loss function, centered on wavelet domain analysis. Images are transformed using DWT/SWT, and a composite loss function with frequency-specific constraints is built: MSE optimizes low-frequency components to ensure structural fidelity of key diagnostic features, while LPIPS refines high-frequency components to enhance detail perception without artifacts, preserving structural integrity and detail semantic authenticity. Experiments show our wavelet-domain frequency-aware loss outperforms traditional methods in objective metrics and downstream task accuracy. It helps to suppress artifacts, retain critical diagnostic information, and boost the clinical utility of SR results for auxiliary diagnosis.