面向冷冻电子断层成像联合复原的监督式多任务框架：由生成式物理仿真驱动
A Supervised Multi-task Framework for Joint cryo-ET Restoration Enabled by Generative Physical Simulation
用真实结构（EMDB）驱动的物理成像仿真生成带干净真值的断层图，把「缺真值」的低层复原任务转成全监督多任务学习；这套仿真-监督范式可迁移到其他缺乏 GT 的退化成像复原（如低剂量 CT、显微去噪）。
原文依据 an imaging simulation pipeline that captures authentic noise distributions and incorporates isotropic structural priors by simulating tomograms from real EMDB structures
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论冷冻电子断层成像（cryo-ET）能够原位可视化细胞超微结构，但由于剂量限制和倾角受限，重建结果受极低信噪比和缺失楔伪影的严重退化。现有基于学习的方法还受制于不准确的噪声建模和可靠真值的缺乏，限制了复原质量。我们提出 cryoDeRec，一个以全监督方式联合执行去噪与缺失楔恢复的多任务框架。我们的关键贡献是一种双目标训练策略，利用合成的含噪断层图及其对应的干净断层图，在恢复缺失信息的同时提升结构保真度。为支撑训练，我们引入一条成像仿真流程，通过从真实 EMDB 结构仿真断层图来刻画真实噪声分布并引入各向同性结构先验。在四个真实 cryo-ET 数据集和两个极低信噪比仿真数据集（均通过 WBP 重建）上的实验表明，cryoDeRec 无需预处理即可直接从原始输入复原高质量断层图，持续优于此前的最先进方法。我们的结果表明，在能够刻画真实噪声与结构的综合仿真数据集上训练，可使模型有效泛化到真实 cryo-ET 断层图。
英文原文摘要 1392 字符
Cryo-electron tomography (cryo-ET) enables in-situ visualization of cellular ultrastructure, but reconstructions are severely degraded by extremely low SNR and missing-wedge artifacts due to dose limits and restricted tilt angles. Existing learning-based approaches are further constrained by inaccurate noise modeling and the lack of reliable ground truth, limiting restoration quality. We propose cryoDeRec, a multi-task framework that jointly performs denoising and missing wedge recovery in a fully supervised manner. Our key contribution is a dual-objective training strategy that leverages synthetic noisy tomograms and their corresponding clean tomograms, encouraging structural fidelity while recovering missing information. To support training, we introduce an imaging simulation pipeline that captures authentic noise distributions and incorporates isotropic structural priors by simulating tomograms from real EMDB structures. Experiments on four realistic cryo-ET datasets and two extremely low-SNR simulated datasets (all reconstructed via WBP) show that cryoDeRec restores high-quality tomograms directly from raw inputs without preprocessing, consistently outperforming prior state of the art. Our findings show that training on a comprehensive simulated dataset, which captures realistic noise and structure, enables models to generalize effectively to real cryo-ET tomograms.