Skip to main content
QUICK REVIEW

[论文解读] LAW & ORDER: Adaptive Spatial Weighting for Medical Diffusion and Segmentation

Anugunj Naman, Ayushman Singh|arXiv (Cornell University)|Mar 5, 2026
Advanced Neural Network ApplicationsComputer Science被引用 0
一句话总结

本文提出 LAW 与 ORDER,自适应空间加权用于扩散与分割,提升扩散保真度与下游分割效果,设计紧凑高效。

ABSTRACT

Medical image analysis relies on accurate segmentation, and benefits from controllable synthesis (of new training images). Yet both tasks of the cyclical pipeline face spatial imbalance: lesions occupy small regions against vast backgrounds. In particular, diffusion models have been shown to drift from prescribed lesion layouts, while efficient segmenters struggle on spatially uncertain regions. Adaptive spatial weighting addresses this by learning where to allocate computational resources. This paper introduces a pair of network adapters: 1) Learnable Adaptive Weighter (LAW) which predicts per-pixel loss modulation from features and masks for diffusion training, stabilized via a mix of normalization, clamping, and regularization to prevent degenerate solutions; and 2) Optimal Region Detection with Efficient Resolution (ORDER) which applies selective bidirectional skip attention at late decoder stages for efficient segmentation. Experiments on polyp and kidney tumor datasets demonstrate that LAW achieves 20% FID generative improvement over a uniform baseline (52.28 vs. 65.60), with synthetic data then improving downstream segmentation by 4.9% Dice coefficient (83.2% vs. 78.3%). ORDER reaches 6.0% Dice improvement on MK-UNet (81.3% vs. 75.3%) with 0.56 GFLOPs and just 42K parameters, remaining 730x smaller than the standard nnUNet.

研究动机与目标

  • 解决医学影像中病变在大背景中占很小区域造成的空间不均衡问题。
  • 开发可应用于扩散式合成与分割的统一自适应加权框架。
  • 通过区域感知的资源分配实现更高的生成保真度和更好的下游分割。
  • 保持在实用医学影像流程中的轻量化架构下的高效性。

提出的方法

  • LAW(可学习自适应加权器)从中间扩散特征与条件掩模预测逐像素的损失调制,并通过归一化、钳制和 Dice 正则化来稳定训练。
  • ORDER(高效分辨率的最佳区域检测)在轻量级 MK-UNet 骨干的后期解码阶段加入选择性双向跳跃注意力,使用共享相似性矩阵以降低注意力成本。
  • LAW 基于带掩模条件的扩散与比率先验,扩展一个 delta 映射以产生自适应权重,并通过最终归一化来界定权重。
  • ORDER 保留 MK-UNet 的多核编码器,并在晚期阶段的双向注意机制中引入置信门控以缩放残差更新。
  • 训练包含 LAW 加权的扩散损失、教师监督以及可选的蒸馏损失;分割使用 BCE+Dice,结合选择性注意与融合。
Figure 7 : Qualitative segmentation results from ORDER on challenging polyp cases. Each row shows: input image, ground truth mask, and predicted segmentation. ORDER produces accurate boundaries even on difficult cases with varying polyp sizes and shapes, demonstrating the effectiveness of selective
Figure 7 : Qualitative segmentation results from ORDER on challenging polyp cases. Each row shows: input image, ground truth mask, and predicted segmentation. ORDER produces accurate boundaries even on difficult cases with varying polyp sizes and shapes, demonstrating the effectiveness of selective

实验结果

研究问题

  • RQ1自适应空间加权如何在医学图像合成的扩散模型中改善病变聚焦的去噪?
  • RQ2学习得到的、基于特征的加权是否可以优于固定先验以强调区域?
  • RQ3在轻量化分割模型中,选择性双向注意力是否能在不过度计算的情况下提升对空间不确定区域的分割?
  • RQ4在用合成数据训练时,自适应加权对下游分割性能有何影响?
  • RQ5在超紧凑的分割骨干中应用自适应加权时,是否存在实际的效率-精度平衡?

主要发现

  • LAW 相对于均匀基线在 Polyps 数据集上实现了 20% 的 FID 提升(65.60 → 52.28)。
  • LAW 在以 LAW 生成的数据训练下实现下游分割 Dice 提升 6%(78.3 → 83.2)。
  • ORDER 相较于 MK-UNet 提升 Dice 6.0%(81.3% vs 75.3%),仅 42K 参数。
  • ORDER 的计算量为 0.56 GFLOPs、参数 0.56G,约比 nnUNet 小 730 倍,同时 Dice 提升到 81.3%、IoU 提升至 81.7%。
  • LAW 与 ORDER 在 Polyps 和 KiTS19 数据集上均显示稳定增益,验证自适应空间加权在生成任务与判别任务中的共通原理。
Figure 8: Visualization of learned attention maps from LAW during training. Top row : Conditioning masks. Bottom row : Learned attention maps (delta maps). At step 0, attention is uniform. By step 250, the model begins focusing on lesion regions. At step 500, attention strongly emphasizes lesion bou
Figure 8: Visualization of learned attention maps from LAW during training. Top row : Conditioning masks. Bottom row : Learned attention maps (delta maps). At step 0, attention is uniform. By step 250, the model begins focusing on lesion regions. At step 500, attention strongly emphasizes lesion bou

阅读本论文

在 Nubint 阅读器中打开原文,边读边就任意段落提问。

在 Nubint 中阅读本论文

本解读由 AI 根据论文原文生成,引用前请核对原文。