Fraunhofer Institute for Digital Medicine MEVIS, Max-von-Laue-Str. 2 28359, Bremen, Germany.
Department of Diagnostic and Interventional Radiology, University Hospital Würzburg, Oberdürrbacher Str. 6 97080, Würzburg, Germany.
Eur J Radiol. 2024 Jul;176:111534. doi: 10.1016/j.ejrad.2024.111534. Epub 2024 May 25.
Radiological reporting is transitioning to quantitative analysis, requiring large-scale multi-center validation of biomarkers. A major prerequisite and bottleneck for this task is the voxelwise annotation of image data, which is time-consuming for large cohorts. In this study, we propose an iterative training workflow to support and facilitate such segmentation tasks, specifically for high-resolution thoracic CT data.
Our study included 132 thoracic CT scans from clinical practice, annotated by 13 radiologists. In three iterative training experiments, we aimed to improve and accelerate segmentation of the heart and mediastinum. Each experiment started with manual segmentation of 5-25 CT scans, which served as training data for a nnU-Net. Further iterations incorporated AI pre-segmentation and human correction to improve accuracy, accelerate the annotation process, and reduce human involvement over time.
Results showed consistent improvement in AI model quality with each iteration. Resampled datasets improved the Dice similarity coefficients for both the heart (DCS 0.91 [0.88; 0.92]) and the mediastinum (DCS 0.95 [0.94; 0.95]). Our AI models reduced human interaction time by 50 % for heart and 70 % for mediastinum segmentation in the most potent iteration. A model trained on only five datasets achieved satisfactory results (DCS > 0.90).
The iterative training workflow provides an efficient method for training AI-based segmentation models in multi-center studies, improving accuracy over time and simultaneously reducing human intervention. Future work will explore the use of fewer initial datasets and additional pre-processing methods to enhance model quality.
放射学报告正在向定量分析转变，这需要对生物标志物进行大规模的多中心验证。这项任务的一个主要前提和瓶颈是对图像数据进行体素标注，对于大样本量来说这是一项耗时的工作。在这项研究中，我们提出了一种迭代式的训练工作流程，以支持和促进这种分割任务，特别是针对高分辨率的胸部 CT 数据。
我们的研究包括 132 例来自临床实践的胸部 CT 扫描，由 13 名放射科医生进行标注。在三个迭代式的训练实验中，我们旨在改善和加速心脏和纵隔的分割。每个实验都从手动分割 5-25 例 CT 扫描开始，这些扫描作为 nnU-Net 的训练数据。进一步的迭代纳入了 AI 预分割和人工修正，以提高准确性、加速标注过程，并随着时间的推移减少人工干预。
结果显示，随着每次迭代，AI 模型的质量都有一致的提高。重新采样数据集提高了心脏（DCS 0.91[0.88;0.92]）和纵隔（DCS 0.95[0.94;0.95]）的 Dice 相似系数。在最有效的迭代中，我们的 AI 模型将心脏和纵隔分割的人工交互时间分别减少了 50%和 70%。仅使用五个数据集训练的模型就能取得令人满意的结果（DCS>0.90）。
迭代式的训练工作流程为多中心研究中基于 AI 的分割模型的训练提供了一种高效的方法，随着时间的推移提高了准确性，同时减少了人工干预。未来的工作将探索使用更少的初始数据集和额外的预处理方法来提高模型质量。
