Source-linked AI summary
Coronary Artery Centerline Extraction in Cardiac CT Angiography Using a CNN-Based Orientation Classifier
Jelmer M. Wolterink, Robbert W. van Hamersvelt, Max A. Viergever, Tim Leiner, Ivana Išgum
TL;DR
Coronary centerline extraction in CCTA is needed for stenosis and plaque assessment, but manual tracing is impractical. The paper uses a 3D dilated CNN to predict vessel direction and radius from local patches and tracks from one seed point, achieving accurate centerlines and broad automatic segment coverage.
Problem
Manual coronary centerline extraction supports CCTA analysis but is tedious and time-consuming, making it infeasible in clinical practice.
Method
A 3D dilated CNN predicts coronary centerline direction and lumen radius from a local image patch, guiding two-way tracking from a single manual or automatic seed point.
Results
The method achieved high-speed, high-accuracy centerline extraction and extracted on average 92% of clinically relevant coronary artery segments automatically.
Takeaways & Limitations
The trained method supports fast automatic or interactive extraction of individual coronary centerlines and full coronary artery trees from CCTA images.
Takeaways & Limitations
Centerline extraction was most affected by step-and-shoot artifacts and sometimes failed to reach ostia with severe calcifications, stenosis, or sharply angled branching.
Abstract
from arXiv · showhide
Coronary artery centerline extraction in cardiac CT angiography (CCTA) images is a prerequisite for evaluation of stenoses and atherosclerotic plaque. We propose an algorithm that extracts coronary artery centerlines in CCTA using a convolutional neural network (CNN). A 3D dilated CNN is trained to predict the most likely direction and radius of an artery at any given point in a CCTA image based on a local image patch. Starting from a single seed point placed manually or automatically anywhere in a coronary artery, a tracker follows the vessel centerline in two directions using the predictions of the CNN. Tracking is terminated when no direction can be identified with high certainty. The CNN was trained using 32 manually annotated centerlines in a training set consisting of 8 CCTA images provided in the MICCAI 2008 Coronary Artery Tracking Challenge (CAT08). Evaluation using 24 test images of the CAT08 challenge showed that extracted centerlines had an average overlap of 93.7% with 96 manually annotated reference centerlines. Extracted centerline points were highly accurate, with an average distance of 0.21 mm to reference centerline points. In a second test set consisting of 50 CCTA scans, 5,448 markers in the coronary arteries were used as seed points to extract single centerlines. This showed strong correspondence between extracted centerlines and manually placed markers. In a third test set containing 36 CCTA scans, fully automatic seeding and centerline extraction led to extraction of on average 92% of clinically relevant coronary artery segments. The proposed method is able to accurately and efficiently determine the direction and radius of coronary arteries. The method can be trained with limited training data, and once trained allows fast automatic or interactive extraction of coronary artery trees from CCTA images.
1. Introduction
Coronary centerline extraction is clinically useful but manual tracing is too time-consuming, motivating automated methods. The paper proposes learning vessel direction and radius directly from CCTA images to avoid hand-crafted intermediate representations.
- Manual coronary centerline extraction is tedious and time-consuming, making it infeasible for clinical practice.
- Existing approaches commonly require filters or models to determine vessel location, orientation, and radius.
- Hand-crafted vessel filters and models may fail when coronary appearance assumptions do not hold, including at branches or non-coronary structures.
- Some learned methods still require intermediate feature representations and evaluations at multiple voxel locations to estimate orientation at one point.
- The proposed CNN learns coronary centerline direction and lumen radius directly from image data without intermediate hand-crafted vesselness representations.
- The method uses direction and radius predictions to extract centerlines from a single seed point placed manually or automatically.
2. Data
The study evaluates the method across publicly available challenge data, institutional scans, and scans from additional CT scanners. These datasets include manual vessel annotations, seed markers, and coronary ostium annotations for different evaluation settings.
- The study uses CAT08, UMCU, and orCaScore CCTA scans acquired across multiple CT scanners.
- CAT08 contains 32 CCTA images divided into training and test sets according to subjective image quality and coronary calcium burden.
- The UMCU scans were consecutively acquired at UMC Utrecht using a 128-detector-row Philips Brilliance iCT scanner.
- In the UMCU dataset, experts placed 5,448 coronary artery markers approximately 10 mm apart across visible major arteries and branches.
- The orCaScore dataset comprises 36 scans acquired equally on GE Lightspeed VCT and Toshiba Aquilion ONE scanners.
3. Method
The method uses a 3D CNN to estimate coronary-vessel direction and radius from local image patches, then supports iterative and fully automatic centerline extraction. Dilated convolutions, direction classification, radius regression, augmented training samples, and CNN-based seeding form the core pipeline.
- CNN architecture: The CNN maps a 3D image patch to a posterior distribution over tracking directions and a regression estimate of vessel radius.The input patch is centered at location x; direction is classified over D, while radius is predicted as a scalar.
- CNN architecture: 27 trainable parameters per dilated kernel expand the receptive field from 3 to 7 to 15 voxels while parameter growth remains linear.Stacked 3 × 3 × 3 kernels use increasing dilation levels to aggregate multiscale information efficiently.
- CNN architecture: Direction classification represents possible directions on a sphere and preserves multimodal posterior predictions for tracking.The tracker selects the highest-probability direction, while the architecture combines |D| classification outputs with one radius-regression output.
- Training strategy: Training samples pair image patches with two opposite reference directions and a reference radius derived from annotated centerlines.Samples are augmented with off-centerline translations and random rotations to improve robustness to tracking deviations and vessel orientation.
- Fully automatic coronary tree extraction: Fully automatic extraction uses one CNN to identify tracker seed points and a second CNN to locate coronary ostia for identifying successfully tracked centerlines.The extension removes the need to place every seed point manually when extracting coronary trees across many scans.
4. Experiments and results
Experiments evaluated centerline extraction accuracy, robustness, radius prediction, and fully automatic coronary-tree recovery across multiple CCTA datasets. The method achieved high overlap and sub-voxel centerline accuracy, while performance varied with image quality, artifacts, anatomy, receptive-field size, and data augmentation.
- Efficiency: 0.4 ± 0.1 s was the average processing time for extracting a single vessel; fully automatic coronary-tree extraction took around 20 s per patient.Automatic seed and ostium identification each required around 1 s per patient.
- Orientation resolution: 500, 1000, and 2000 orientations produced the best inside-accuracy results, so the method used |D| = 500 thereafter.Increasing |D| produced finer-grained but noisier CNN responses.
- Centerline extraction accuracy: 93.7% OV, 81.5% OF, 97.0% OT, and 0.21 mm AI were obtained on 24 CAT08 test images containing 96 coronary arteries.These results corresponded to an average CAT08 rank of 5.76 and third place among the compared methods.
- Centerline extraction robustness: Step-and-shoot artifacts, severe calcifications or stenoses, and sharply angled branches could prevent centerlines from reaching the ostium.Additional seed points may overcome some of these limitations.
- Fully automatic centerline extraction: 92% of clinically relevant coronary-tree segments visualized in 36 scans were identified by fully automatic extraction, with all segments identified in 16 scans.Automatic ostium detection succeeded in all 36 images with an average error of 1.8 ± 1.0 mm.
- Radius prediction: Radius errors in CAT08 had a systematic bias of -0.01 mm and 95% limits of agreement from −0.39 mm to 0.41 mm, whereas UMCU showed 0.31 mm overestimation.The difference was partly attributed to differing annotation protocols and reference-radius distributions.
- Effect of data augmentation: Translation augmentation improved both overlap and accuracy, and enabled recovery from off-centerline seeds that otherwise caused extreme sensitivity to seed placement.Rotation augmentation mostly improved overlap; using both augmentation types substantially improved both metrics.
5. Discussion
The CNN directly predicts coronary direction and lumen radius, enabling accurate, fast tracking from sparse annotations and single seed points. Fully automatic extraction remains constrained by venous confusion and the need for representative annotated training data.
- Method: The CNN directly predicts coronary centerline direction and lumen radius from CCTA image patches without hand-crafted filters or features.The architecture uses stacked dilated convolutions to enlarge the receptive field without reducing feature resolution.
- Single-vessel tracking: 0.21 mm average distance to manually annotated reference centerlines was achieved in CAT08 evaluation, with accuracy ranking third among 25 publicly reported results.The tracker required only one seed point, which could be placed anywhere in the coronary artery, and automatically identified endpoints from CNN prediction entropy.
- Training requirements: The supervised method requires representative manually annotated training scans, although exhaustive annotation of every coronary artery was unnecessary in the reported training setup.The eight training images contained annotations for only four coronary artery centerlines rather than the full coronary tree.
- Generalization: Rotation and off-centerline augmentation improved generalization by teaching rotational invariance and corrections toward the centerline and endpoints.Systematic differences in contrast attenuation can still cause occasional confusion between arteries and veins.
- Automatic tree extraction: The method missed 1.17 coronary segments per CCTA scan during fully automatic extraction, compared with 1.25 reported for a similar previous analysis.The automatic scheme merged extracted centerlines using a bottom-up approach.
- Single-vessel tracking: 0.4 s was sufficient for interactive extraction of a coronary artery from one seed point.The seed point only needed to lie within the artery of interest rather than at a proximal or distal endpoint.
6. Conclusions
The paper presents a deep learning method for coronary artery centerline extraction. Its CNN simultaneously determines centerline direction and coronary lumen radius with high speed and accuracy.
- A CNN-based method simultaneously determines coronary artery centerline direction and coronary lumen radius with high speed and accuracy.