Pseudo-healthy image synthesis aims to generate subject-specific, pathology-free images from pathological scans. Such images can be helpful in certain tasks, such as anomaly detection and understanding changes induced by pathology and disease. A participant cannot be "healthy" and "unhealthy" at the same time, and thus, directly obtaining pathological and healthy paired images of the same individual to train and evaluate supervised learning algorithms is infeasible. In addition, simultaneously meeting the requirements of subjects' "identity" preservation and pathology restoration performance is frequently difficult for existing unsupervised learning methods, especially for large or information-free pathological regions, such as postoperative cavities. In this study, we propose a novel pseudo-healthy synthesis framework that combines low-resolution residual decoupling with an edge-prior-guided super-resolution reconstruction module. We named this framework LRD-ESR-Net. In particular, by using a coarse-to-fine synthesis pipeline, the residual decoupling network first decouples information-rich tumor tissues or information-free resection cavities from healthy brain tissues in low-resolution pathological magnetic resonance images. Then, a residual-shifting diffusion network with Canny edge maps is employed to reconstruct low-resolution pseudo-healthy images to their original resolutions. We evaluate the proposed framework on one in-house brain dataset, two public brain datasets, and one public liver dataset, and validate its effectiveness on low-contrast lesion segmentation and pre-/postoperative brain tumor MRI registration. Results show that LRD-ESR-Net consistently outperforms state-of-the-art methods in pseudo-healthy image quality, anatomical preservation, and downstream task performance, demonstrating strong robustness and generalization across organs, modalities, and lesion types.

The growing demand for efficient deep learning has positioned dataset distillation as a pivotal technique for compressing training dataset while preserving model performance. However, existing inner-loop optimization methods for dataset distillation typically rely on random truncation strategies, which lack flexibility and often yield suboptimal results. In this work, we observe that neural networks exhibit distinct learning dynamics across different training stages—early, middle, and late—making random truncation ineffective. To address this limitation, we propose Automatic Truncated Backpropagation Through Time (AT-BPTT), a novel framework that dynamically adapts both truncation positions and window sizes according to intrinsic gradient behavior. AT-BPTT introduces three key components: (1) a probabilistic mechanism for stage-aware timestep selection, (2) an adaptive window sizing strategy based on gradient variation, and (3) a low-rank Hessian approximation to reduce computational overhead. Extensive experiments on CIFAR-10, CIFAR-100, Tiny-ImageNet, and ImageNet-1K show that AT-BPTT achieves state-of-the-art performance, improving accuracy by an average of 6.16\% over baseline methods. Moreover, our approach accelerates inner-loop optimization by 3.9 × while saving 63\% memory cost.