Spatial-SAM：基于 SDF 记忆与半监督学习的空间一致三维电子显微镜分割
Spatial-SAM: Spatially Consistent 3D Electron Microscopy Segmentation with SDF Memory and Semi-Supervised Learning
用 3D U-Net 预计算的有符号距离场（SDF）替代 SAM2 的记忆库，为视频/体数据分割提供几何先验以维持跨切片空间一致性；配合双轨伪标签迭代，可在极少标注下把 SAM2 的少样本能力迁移到 3D 体数据，这套「SDF 记忆 + 伪标签自训练」思路可复用到其他层序切片数据（CT、MRI、显微）的 3D 分割。
原文依据 replaces SAM2's memory with SDF representations precomputed by a 3D U-Net
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论基于 Segment Anything Model（SAM）的方法在生物医学图像分割中已展现出强大潜力。然而，这些方法在三维电子显微镜（3D-EM）数据中往往难以保持空间一致性，并且仍然需要大量人工标注。我们提出 Spatial-SAM，一个面向高精度 3D-EM 分割的空间一致且标注高效的框架。它引入三维有符号距离场（SDF）记忆机制，用由 3D U-Net 预计算的 SDF 表示替代 SAM2 的记忆，从而提供更丰富的几何信息并提升空间一致性。它还将 SAM2 的少样本能力与双轨伪标签迭代优化策略相结合，以从极少量标注中学习大规模 3D-EM 分割。实验表明，Spatial-SAM 在多个 3D-EM 基准上显著优于现有半监督方法，并与最优的全监督方法性能相当，在保持空间一致性的同时降低了标注成本。代码见 https://github.com/Giluir/Spatial-SAM。
英文原文摘要 1090 字符
Segment Anything Model (SAM)-based approaches have shown strong potential for biomedical image segmentation. However, these methods often struggle to preserve spatial consistency in 3D electron microscopy (3D-EM) data and still require extensive manual annotation. We propose Spatial-SAM, a spatially consistent and annotation-efficient framework for high-precision 3D-EM segmentation. It introduces a 3D Signed Distance Field (SDF) memory mechanism that replaces SAM2's memory with SDF representations precomputed by a 3D U-Net, providing richer geometric information and improving spatial consistency. It also combines SAM2's few-shot capability with a dual-track pseudo-label iterative optimization strategy to learn large-scale 3D-EM segmentation from minimal annotations. Experiments show Spatial-SAM significantly outperforms existing semi-supervised methods and performs comparably to state-of-the-art fully supervised approaches on multiple 3D-EM benchmarks, reducing annotation costs while preserving spatial consistency. Code is available at https://github.com/Giluir/Spatial-SAM.