Intentional Degradation for Stabilizing Residual Learning in Coarse-to-Fine Medical Image Refinement.

Background: Two-stage coarse-to-fine architectures are commonly used in medical image synthesis and refinement. However, when the distributional gap between coarse inputs and high-resolution targets is large, the second-stage generator may produce unstable or weakly conditioned refinements rather than operating as a true residual corrector. Methods: We propose intentional degradation, a target-side distribution-alignment strategy that deliberately degrades high-resolution targets toward the coarse prediction space before residual learning. The degradation profile is guided by measured differences in contrast, saturation, and edge energy. We further evaluate refinement quality using two residual-space metrics: Var(Δ), the variance of the predicted residual (target minus input), and Δ-SSIM, the structural similarity between the predicted and true residual maps; both are designed to reveal refinement instability that conventional image-level metrics (e.g., SSIM computed on the full image) may not capture. We evaluated the approach in three medical imaging domains: wound-healing photography, retinal fundus imaging, and dermoscopy. Results: Across all three domains, intentional degradation consistently improved Δ-SSIM and normalized residual variance recovery relative to the baseline, and predicted residuals showed substantially stronger directional correspondence with the true residual (assessed via sign-agreement and cosine-similarity metrics) than the baseline, which converged toward directionally random predictions. These findings suggest that residual-space and direction-sensitive metrics can reveal refinement instability that may not be captured by conventional image-level metrics alone. Conclusions: Intentional degradation provides a simple, architecture-agnostic-by-design strategy for stabilizing residual learning in coarse-to-fine medical image refinement, validated within a single ConvLSTM-based refinement framework; extension to other architectures remains for future validation. Rather than introducing a new network architecture, the method modifies the target-side training distribution to reduce the mismatch between coarse inputs and high-resolution targets.