eess.IVAug 25, 2026

Unified-protocol voxel-level pulmonary embolism annotations for three public CT angiography datasets

Authors: Qihang Sun, Zhongxiao Liu, Bailiang Jian, Shenman Qiu, Jingyuan Wang, Lei Zhang, Lixiang Xie, Jiazhen Pan, +1 more

Organizations: Technical University of Munich (TUM), Munich, Germany · Munich Center for Machine Learning (MCML), Munich, Germany · Department of Radiology, The Affiliated Hospital of Xuzhou Medical University, Xuzhou, China · Department of Radiology, The Third Affiliated Hospital of Soochow University, Changzhou, China · Department of Radiology, The Affiliated Taizhou People’s Hospital of Nanjing Medical University, Taizhou, China

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Reliable clot-volume quantification and subsequent risk assessment in pulmonary embolism depend on precise segmentation of emboli on computed tomography pulmonary angiography. Deep learning models for this task must be trained on accurate voxel-level labels. The three public datasets that provide such labels were annotated under different protocols, and some of their studies contain unlabeled emboli or labels that are discontinuous across slices. This Data Descriptor presents voxel-level pulmonary embolism annotations for 149 of the 166 studies in these datasets. A primary rater drew all annotations under a single protocol. A thoracic radiologist with more than 20 years of experience reviewed and revised them. Three raters at three different centers independently annotated a subset of 15 studies. The subset was selected by source dataset and embolus location. Technical validation quantifies volumetric agreement with the source annotations, changes in within-mask attenuation, and inter-rater agreement on the subset. The dataset is intended to allow segmentation models to be developed and compared under a common reference standard.

Figures & tables

CharacteristicCAD-PE (n = 91)FUMPE (n = 35)READ (n = 40)
Sex, female17 (43) †18 (51)27 (68)
Sex, male23 (57) †17 (49)13 (32)
Age, y65.0 ±\pm 18.0 †56.6 ±\pm 19.7 †55.5 ±\pm 19.2
CountrySpainIranBrazil
InclusionPE-positive onlyPE-positive only (33/35 confirmed)PE-positive only
ScannerSiemens Somatom SensationMixedToshiba Aquilion; GE Revolution
Table 1: Patient demographics, acquisition characteristics and source reference standards of the three public CT pulmonary angiography datasets
Figure 1: Composition and content of the harmonized collection. (a) Screening of the 166 studies in the three source datasets. Of these, 149 studies were included (CAD-PE, n = 76; FUMPE, n = 33; READ, n = 40). The identifiers of the excluded studies are shown. (b) Mean CT attenuation of the pulmonary artery trunk per dataset, in Hounsfield units. (c) Per-study distribution of the most proximal level of clot involvement. (d) Embolus density of the harmonized masks after structure-wise mapping into a common reference space, per dataset and pooled (ALL): anterior view of the lung volume (top row), medial mediastinal surface view (middle row) and pulmonary arterial surface (bottom row). Colors show relative embolus density on a nonlinear scale ( γ=0.25\gamma=0.25 ), with separate scales for the lung lobes and the pulmonary artery, each shared across datasets. R and L denote the right and left lung.
Figure 2: Representative differences between the source and the harmonized annotations. Red boxes mark regions included in the source annotation but excluded under the harmonized label definition; green boxes mark internal voids in the source foreground; red arrows mark emboli annotated under the harmonized definition but absent from the source annotation. Study 001: source foreground extending into the opacified pulmonary artery. Study e0678: red box, source foreground extending into the pulmonary artery; green box, void within the annotated region. Study 007: source foreground extending into a pulmonary vein. Study Patient16: red arrow, embolus present in the harmonized annotation only.
Figure 3: Per-lesion volume, per-study embolus count, difference categories and mask attenuation under the source and the harmonized annotations, by dataset. (a) Kernel density estimates of per-lesion volume on a logarithmic scale for the source (dashed) and harmonized (solid) annotations, colored by dataset, with triangles at the medians. (b) Per-study lesion counts as box plots; light fill, source; dark fill, harmonized. In all box plots the box spans the interquartile range, the line marks the median, whiskers extend to 1.5 times the interquartile range and points beyond are drawn individually. (c) Radar plot of the predefined difference categories recorded by the two reviewing raters, as a percentage of studies per dataset. FP PA, FP PV, FP Lung and FP Bronchi denote source foreground extending onto pulmonary artery, pulmonary vein, lung parenchyma or bronchi; FN denotes an embolus present in the harmonized annotation only; Error denotes anatomically implausible foreground. (d) Per-lesion mask attenuation, mean and standard deviation in Hounsfield units, under the source and harmonized annotations, shown as combined box, strip and half-violin plots.
Figure 4: Source and harmonized masks with the multi-rater consensus on representative studies. For each study the same embolus is shown in the axial and coronal planes; the leftmost image of each plane shows the whole slice and the two adjacent panels, Zoom A and Zoom B, magnify the two regions outlined by the red dashed boxes labeled A and B on the full image. Overlays: source annotation in red, Annotation 1 in green, and the STAPLE consensus of Annotations 2 to 4 in cyan. Ax denotes axial and Cor denotes coronal.
Figure 5: Pairwise agreement between the four harmonized delineations and the source annotation on the 15-study multi-rater subset. Each panel is a 5 ×\times 5 matrix over Annotations 1 to 4 and the source annotation. (a–f) Agreement across all 15 studies, as lower-triangular pairwise means: (a) Dice similarity coefficient, (b) average symmetric surface distance in mm, (c) normalized surface distance at a 1 mm tolerance, and (d–f) lesion-level F1F_{1} at overlap criteria of 1 pixel, 10% and 20%. (g–l) The same six measures stratified by embolus volume: studies were sorted by total volume on the STAPLE consensus of all four annotations and split into a smaller-volume half of 7 studies, shown in the lower triangle, and a larger-volume half of 8 studies, shown in the upper triangle, at a cut point of 2 mL: (g) Dice similarity coefficient, (h) average symmetric surface distance in mm, (i) normalized surface distance at 1 mm, and (j–l) lesion-level F1F_{1} at 1 pixel, 10% and 20%. Color bars encode the scale of each measure; the dashed diagonal in (g–l) separates the two volume subgroups.

Explore similar work

CardsList

  1. A Patient-Specific Pulmonary Arterial Tree Digital Twin to Extract Pulmonary Embolism Biomarkers

    May 27, 2026Morgane des Ligneris, Nathan Painchaud, Allan Serva +4Coronary AngiographyCardiovascular Risk Prediction

    Univ Lyon, INSA-Lyon, Universit´e Claude Bernard Lyon 1, UJM-Saint Etienne, CNRS, Inserm, CREATIS UMR 5220, U1294, F-69621, Lyon, France · Department of Pneumology, CHU Saint-Etienne, UJM Saint-Etienne, France and Universit´e Jean Monnet Saint- ´Etienne, CHU Saint- ´Etienne, Mines Saint- ´Etienne, INSERM, SAINBIOSE U1059, CIC 1408, D´epartement de M´edecine Vasculaire et Th´erapeutique, F-CRIN INNOVTE network, all in F-42055, Saint- ´Etienne, France · Department of Radiology, CHU Saint-Etienne, UJM Saint-Etienne, Saint-Etienne, France +1

  2. Pulmonary Embolism Risk Stratification from CTPA and Medical Records: Vascular Graphs Are Not All You Need

    Jun 24, 2026Nathan Painchaud, Tristan Habémont, Morgane des Ligneris +6Cardiovascular Risk PredictionCoronary Angiography

    INSA-Lyon, Université Lyon 1, UJM-Saint Etienne, CNRS, Inserm, CREATIS UMR 5220, U1294, F-69621, Lyon, France · Université Jean Monnet Saint-Étienne, CHU Saint-Étienne, Mines Saint-Étienne, INSERM, SAINBIOSE U1059, CIC 1408, Département de Médecine Vasculaire et Thérapeutique, F-CRIN INNOVTE network, all in F-42055, Saint-Étienne, France · Department of Pneumology, CHU Saint-Etienne, UJM Saint-Etienne, France +3

  3. Efficient Multimodal Clinical Question Answering for Pulmonary Embolism Risk Assessment

    Jun 21, 2026Xiangyuan Xue, Yang Yu, Yan Gao +5Multimodal Clinical DataElectronic Health Records

    University of Auckland, Auckland, New Zealand · University of Cambridge, Cambridge, United Kingdom · University of Adelaide, Adelaide, Australia +1