中文

基于维数缩减与最近邻居提高医学图像分割中的离群检测

计算机视觉与模式识别 2024-10-03 v3 机器学习

摘要

临床部署的深度学习分割模型已知在数据超出训练分布时会失效。尽管临床医生会审查这些分割结果,但这些模型在大多数情况下仍能表现良好,这可能会加剧自动化偏见。因此,在推断时检测离群分布图像至关重要,以警示临床医生该模型可能已失败。本工作将巴马兹距离 (MD) 应用于四个用于分割肝脏的 Swin UNETR 和 nnU-net 模型的瓶颈特征上,这些模型在 T1 加权磁共振成像和 CT 图像上进行。通过对瓶颈特征进行维数缩减(采用主成分分析或统一流形近似和投影),能够以高性能和最小计算负担检测模型失败的图像。此外,本工作探索了 MD 的非参数替代方案,即 k 近邻距离 (KNN)。KNN 在应用于原始和平均池化瓶颈特征时,显著提高了可扩展性和性能。

关键词

引用

@article{arxiv.2408.02761,
  title  = {Dimensionality Reduction and Nearest Neighbors for Improving Out-of-Distribution Detection in Medical Image Segmentation},
  author = {McKell Woodland and Nihil Patel and Austin Castelo and Mais Al Taie and Mohamed Eltaher and Joshua P. Yung and Tucker J. Netherton and Tiffany L. Calderone and Jessica I. Sanchez and Darrel W. Cleere and Ahmed Elsaiey and Nakul Gupta and David Victor and Laura Beretta and Ankit B. Patel and Kristy K. Brock},
  journal= {arXiv preprint arXiv:2408.02761},
  year   = {2024}
}

备注

Accepted for publication at the Journal of Machine Learning for Biomedical Imaging (MELBA) https://melba-journal.org/2024:020. Expansion of "Dimensionality Reduction for Improving Out-of-Distribution Detection in Medical Image Segmentation" arXiv:2308.03723. Code available at https://github.com/mckellwoodland/dimen_reduce_mahal (https://zenodo.org/records/13881989)