Dual-UNet++ Framework for Robust Single-Frame Echocardiographic Dehazing

The Dehazing Echocardiography Challenge 2025 required single-frame denoising, disallowing temporal information. Nevertheless,
we exploited sequence structure during training by applying log-PCA decomposition and using the mid-order reconstruction (up to k = 35
components) as supervised target. This strategy enabled the network to predict temporally consistent reconstructions from individual frames,
transferring temporal coherence into a single-frame setting. High-order components served as structured noise sources, allowing generation of
diverse synthetic corruptions and improving robustness. To enforce anatomical separation, we added an auxiliary segmentation
branch with stochastic threshold sampling of PCA reconstructions, which acted as target augmentation resulting in a stable structure-background
discrimination. Sparse ROI labels were densified via interpolation and further refined using a dedicated auxiliary network.
Our dual UNet++ framework with ResNet backbones achieved a final challenge score of 84.28, with competitive performance across FID, CNR,
gCNR, and Dice. These results demonstrate that temporally informed PCA supervision, structured noise augmentation, and stochastic segmentation targets are effective strategies for robust single-frame echocardiographic dehazing.