文献速递神经介入技术与安全CT

基于二值化深度学习模型在非增强头颅CT上高效自动量化脑出血中线移位

Efficient automated quantification of midline shift in intracerebral hemorrhage using a binarized deep learning model on non-contrast head CT.
PubMed 文献 · Neuroradiology · 2025-11-08

在RSNA 2019出血CT数据集留出测试集5000层上,左右侧Dice分别为0.92和0.91,中线移位平均绝对误差0.09 mm,单层分割中位推理时间5.6 ms。

要点

  • 模型在留出测试集5000层上中线移位平均绝对误差为0.09 mm,左右侧Dice分别为0.92和0.91。
  • 所有物理测量均按原生DICOM间距计算,模型可实时输出中线移位量化结果供决策支持。

核心结论

  • 该资源高效模型能准确量化中线移位,适合急诊神经影像场景,其可重复性和低计算需求使其 ideal 用于低资源环境,为放射科医生提供快速决策支持,并可能 assist 在时间敏感场景下于受限算力预算内准确量化中线移位;需进一步多中心验证。

研究设计与样本

  • 研究提出基于残差U-Net架构的二值化卷积神经网络,使用XNOR二值化和通道缩放,在RSNA 2019出血CT数据集上评估,留出测试集为5000层非增强头颅CT。

关键影像表现/征象

  • 模型分割大脑半球并提取中线特征以量化中线移位,后处理通过边缘检测推导变形中线(dML)并以解剖引导的理想中线(iML)为参考。
  • 中线移位计算为变形中线与理想中线之间的最大水平位移。
  • 在留出测试集5000层上,分割Dice左侧为0.92、右侧为0.91,中线移位平均绝对误差为0.09 mm。

诊断标准与关键数值

  • 中线移位量化指标为变形中线与理想中线之间的最大水平位移,平均绝对误差0.09 mm。
  • 大脑半球分割Dice左侧0.92、右侧0.91。

临床应用与阅片要点

  • 模型在NVIDIA GTX 1650(4 GB)上单层分割中位推理时间为5.6 ms,所有物理测量按原生DICOM间距计算。
  • 作者认为该轻量模型可能 assist 在时间敏感场景下于受限算力预算内准确量化中线移位,适合低资源环境部署。

局限性

  • 原文结论指出需进一步多中心验证(Further multicenter validation is warranted)。

以上解读依据论文摘要(全文未开放获取或许可不允许翻译转载)。

英文原文(摘要)· 点击展开

PURPOSE: Accurate assessment of midline shift (MLS) is critical in the management of patients with acute intracerebral hemorrhage. Manual estimation of MLS by radiologists is time-consuming and subject to inter-observer variability. We present a lightweight, reproducible deep learning model designed for real-time, automated MLS quantification using non-contrast head CT scans.

METHODS: We propose a binarized convolutional neural network based on a residual U-Net architecture that dramatically reduces model parameters (from 31 million to 44 thousand) using XNOR-based binarization and channel-wise scaling. The model segments the cerebral hemispheres and extracts midline features to quantify MLS. Post-processing involves the derivation of the deformed midline (dML) using edge detection and an anatomically guided ideal midline (iML) for reference. MLS is calculated as the maximum horizontal displacement between these lines.

RESULTS: Evaluated on the RSNA 2019 hemorrhage CT dataset (held-out test: 5,000 slices), the model achieved Dice scores of 0.92 (left) and 0.91 (right) and a mean absolute error (MAE) of 0.09 mm for MLS. On an NVIDIA GTX 1650 (4 GB), median inference time was 5.6 ms per slice for the segmentation step; all physical measurements were computed at native DICOM spacing.

CONCLUSION: This clinically-driven, resource-efficient model enables accurate MLS quantification suitable for emergency neuroimaging settings. Its reproducibility and low computational footprint make it ideal for deployment in low-resource environments, providing radiologists with rapid, decision-support insights. This lightweight model may assist accurate MLS quantification in time-sensitive settings while operating within constrained compute budgets. Further multicenter validation is warranted.

来源:Neuroradiology,查看原文。版权归原作者所有。

引用

  1. Nag MK, Sadhu AK, Kumar C, et al. Efficient automated quantification of midline shift in intracerebral hemorrhage using a binarized deep learning model on non-contrast head CT. Neuroradiology. 2025. PMID: 41204958. doi:10.1007/s00234-025-03835-3

供执业医师工作参考,不替代临床判断;数值与分级以指南原文为准。