Data from: Coronary artery segmentation in non-contrast calcium scoring CT images using deep learning

Abstract Precise segmentation of coronary arteries in non-contrast Computed Tomography (CT) scans plays an important role in the assessment of the coronary artery disease, where it is the key component for evaluating the Calcium Score (Agatston et al. 1990). In the paper by Bujny et al. (2024), a deep-learning approach for high-precision segmentation of coronary arteries in non-contrast CT was proposed along with a novel method for generating Ground Truth (GT) test data (test-GT) via manual registration of high-resolution coronary tree models obtained based on contrast CT with the non-contrast CT scans. In this dataset, we present the inferences of the neural network model together with the corresponding test-GT samples, based on 6 CT scans from the openly available OrCaScore dataset (Wolterink et al. 2016). The geometrical models included in the dataset can be used both for inspection of the proposed deep learning model and for testing of new non-contrast coronary vessel segmentation approaches, which is a unique opportunity since, to the best of our knowledge, manual generation of GT for non-contrast coronary artery segmentation was not addressed so far due to very challenging character of this particular segmentation task. Methods Manual Generation of test-GT The geometric models of coronary arteries used for the evaluation of the proposed neural network model were generated according to the manual mesh-to-image registration process as described by Bujny et al. (2024). In this approach, the high-resolution coronary artery masks obtained based on contrast CT scans are manually aligned with the corresponding non-contrast CT images using tools available in the open-source 3D computer graphics software, Blender (https://www.blender.org/). To ease the manual alignment process, specialized add-ons for medical image processing such as Cardiac add-on for Blender of Graylight Imaging (https://graylight-imaging.com/3d-modelling/) can be used, as well. The STL models in this dataset were manually generated by a medical expert with 4 years of experience. Segmentation of Coronary Arteries using a Deep Learning Model For each of the cases presented in this dataset, we run an inference of an nnU-Net (Isensee et al. 2021) model trained according to the process described in our paper (Bujny et al. 2024). Since we use a standard nnU-Net, which utilizes a sliding window approach for processing of the CT scan, the context information within a patch is limited, which can lead to some false-positive detections. To mitigate this problem, we additionally post-process the inferences by eliminating small vessel fragments of less than 50 [mm^3] volume and structures outside of pericardium, which we segment using another nnU-Net model, SegTHOR (Lambert et al. 2020). The resulting geometric models are stored using the STL format and presented as green masks in the HTML reports with an embedded viewer based on the K3D-jupyter library (https://k3d-jupyter.org/). Dataset organization The root folder contains 6 folders whose names correspond to the CT scans from the OrCaScore dataset (Wolterink et al. 2016). In each of the folders, there are the following 3 files available: ‘manualGT.stl’ – high-resolution STL model of coronary arteries obtained via manual alignment of the geometric model segmented in contrast CT with the corresponding non-contrast CT scan. A sample belonging to the test-GT set (Bujny et al. 2024). ‘ML.stl’ – post-processed inference of the nnU-Net ML model in the STL format. ‘report.html’ – interactive HTML report consisting of a manually-aligned test-GT sample (red mask), the ML segmentation based on the non-contrast CT scan (green mask), and selected slices of the non-contrast CT scan. The reports contain the relevant information related to the scanning device and present the main segmentation quality metrics for the ML model inference.摘要 非增强计算机断层扫描（Computed Tomography, CT）中冠状动脉的精准分割在冠状动脉疾病评估中具有重要作用，是评估钙化积分（Calcium Score，Agatston等，1990）的核心环节。在Bujny等（2024）的研究中，研究者提出了一种用于非增强CT冠状动脉高精度分割的深度学习方法，并配套提出了一种全新的真实标注（Ground Truth, GT）测试数据（test-GT）生成方案：通过将基于增强CT获取的高分辨率冠状动脉树模型与非增强CT扫描图像进行手动配准。本数据集基于公开可用的OrCaScore数据集（Wolterink等，2016）中的6例CT扫描数据，提供了该神经网络模型的推理结果与对应的test-GT样本。本数据集包含的几何模型既可用于验证所提出的深度学习模型，也可用于测试新型非增强冠状动脉血管分割方法——这一契机具有独特性，因为据我们所知，由于该细分任务本身极具挑战性，此前尚未有针对非增强冠状动脉分割的手动GT生成方案被提出。 方法 test-GT的手动生成 本研究用于评估所提神经网络模型的冠状动脉几何模型，是依据Bujny等（2024）所述的手动网格-图像配准流程生成的。该方法借助开源3D计算机图形软件Blender（https://www.blender.org/）提供的工具，将基于增强CT扫描获取的高分辨率冠状动脉掩码，手动对齐至对应的非增强CT图像。为简化手动配准流程，还可使用Graylight Imaging开发的Blender心脏插件（https://graylight-imaging.com/3d-modelling/）等医学图像处理专用插件。本数据集内的STL模型均由拥有4年工作经验的医学专家手动生成。 基于深度学习模型的冠状动脉分割 针对本数据集提供的每一例病例，我们运行了依据本团队论文（Bujny等，2024）所述流程训练的nnU-Net（Isensee等，2021）模型进行推理。由于我们采用的标准nnU-Net使用滑动窗口法处理CT扫描数据，其图像块内的上下文信息存在局限，可能导致部分假阳性检测结果。为缓解这一问题，我们额外对推理结果进行后处理：剔除体积小于50 mm³的微小血管片段，以及心包以外的结构——后者通过另一款nnU-Net模型SegTHOR（Lambert等，2020）完成分割。最终得到的几何模型以STL格式存储，并在基于K3D-jupyter库（https://k3d-jupyter.org/）开发的嵌入式查看器生成的HTML报告中以绿色掩码形式展示。 数据集组织结构 根目录包含6个文件夹，文件夹名称与OrCaScore数据集（Wolterink等，2016）中的CT扫描病例一一对应。每个文件夹内均包含以下3个文件： ‘manualGT.stl’：通过将增强CT分割得到的几何模型与对应非增强CT扫描图像手动配准后得到的高分辨率冠状动脉STL模型，属于test-GT测试集样本（Bujny等，2024）。 ‘ML.stl’：经后处理的nnU-Net机器学习模型推理结果，格式为STL。 ‘report.html’：交互式HTML报告，包含手动配准得到的test-GT样本（红色掩码）、基于非增强CT扫描的机器学习分割结果（绿色掩码），以及非增强CT扫描的选定层面。报告中包含了扫描设备的相关信息，并给出了机器学习模型推理结果的主要分割质量评估指标。