Source-linked AI summary
Deep learning analysis of the myocardium in coronary CT angiography for identification of patients with functionally significant coronary artery stenosis
Majd Zreik, Nikolas Lessmann, Robbert W. van Hamersvelt, Jelmer M. Wolterink, Michiel Voskuil, Max A. Viergever, Tim Leiner, Ivana Išgum
TL;DR
Determining whether an intermediate coronary stenosis is functionally significant commonly requires invasive FFR, while CCTA alone has limited specificity. This paper uses deep learning to analyze LV myocardium from a single rest CCTA, achieving moderate patient-classification performance against invasive FFR.
Problem
CCTA has limited specificity for determining functional significance, while invasive FFR is commonly used to assess stenoses and can require unnecessary invasive procedures.
Method
The method automatically segments LV myocardium in a rest CCTA, extracts learned encodings with a CAE, and classifies patients using an SVM.
Results
AUC was 0.74 ± 0.02 for identifying patients with functionally significant stenosis using LV-myocardium features alone.
Takeaways & Limitations
Automatic LV-myocardium analysis from a single rest CCTA can identify patients with functionally significant coronary artery stenosis without assessing coronary anatomy.
Takeaways & Limitations
Segmentation was unsatisfactory where CCTA lacked contrast between the blood pool and LV myocardium or between the right and left ventricles.
Abstract
from arXiv · showhide
In patients with coronary artery stenoses of intermediate severity, the functional significance needs to be determined. Fractional flow reserve (FFR) measurement, performed during invasive coronary angiography (ICA), is most often used in clinical practice. To reduce the number of ICA procedures, we present a method for automatic identification of patients with functionally significant coronary artery stenoses, employing deep learning analysis of the left ventricle (LV) myocardium in rest coronary CT angiography (CCTA). The study includes consecutively acquired CCTA scans of 166 patients with FFR measurements. To identify patients with a functionally significant coronary artery stenosis, analysis is performed in several stages. First, the LV myocardium is segmented using a multiscale convolutional neural network (CNN). To characterize the segmented LV myocardium, it is subsequently encoded using unsupervised convolutional autoencoder (CAE). Thereafter, patients are classified according to the presence of functionally significant stenosis using an SVM classifier based on the extracted and clustered encodings. Quantitative evaluation of LV myocardium segmentation in 20 images resulted in an average Dice coefficient of 0.91 and an average mean absolute distance between the segmented and reference LV boundaries of 0.7 mm. Classification of patients was evaluated in the remaining 126 CCTA scans in 50 10-fold cross-validation experiments and resulted in an area under the receiver operating characteristic curve of 0.74 +- 0.02. At sensitivity levels 0.60, 0.70 and 0.80, the corresponding specificity was 0.77, 0.71 and 0.59, respectively. The results demonstrate that automatic analysis of the LV myocardium in a single CCTA scan acquired at rest, without assessment of the anatomy of the coronary arteries, can be used to identify patients with functionally significant coronary artery stenosis.
1. Introduction
Functionally significant stenoses can cause myocardial ischemia and require severity assessment, but invasive FFR is commonly used and CCTA has limited specificity for functional significance. The proposed approach analyzes LV myocardium in a single rest CCTA instead of relying primarily on coronary anatomy.
- Functionally significant stenoses limit blood flow to the myocardium and can lead to myocardial ischemia.
- Treating stenoses that are not functionally significant can cause more harm than benefit, motivating assessment of their myocardial impact.
- FFR measured during ICA is the reference standard, but commonly used clinical cut-offs range from 0.72 to 0.80.
- CCTA detects coronary artery disease with high sensitivity but limited specificity for functional significance, leading to unnecessary invasive FFR measurements in about 22–52% of patients.
- Existing noninvasive approaches may analyze coronary blood flow or CCTA-derived characteristics, but some require accurate coronary centerlines or additional stress imaging.
- The proposed method identifies patients using only LV myocardium from a single rest CCTA, contrasting with methods focused on coronary anatomy or stenosis.
2. Data
The study used retrospectively collected CCTA scans from 166 patients, with invasive FFR available for 156, and defined reference myocardium through manual annotations. The imaging workflow and annotation excluded specific noncompact myocardial structures.
- The dataset comprised retrospectively collected CCTA scans from 166 patients acquired between 2012 and 2016.The cohort had a mean age of 59.2 ± 9.5 years, and 128 patients were male.
- CCTA scans were acquired with an ECG-triggered step-and-shoot protocol on a 256-detector-row scanner.The protocol used 120 kVp and a tube current between 210 and 300 mAs.
- Invasive FFR measurements were available for 156 patients within one year of CCTA acquisition.The minimum FFR value was used as the patient-level reference when multiple measurements were available.
- Manual LV myocardium annotations were created in 40 randomly selected CCTA scans for segmentation training, validation, and testing.Annotations followed the clinical short-axis workflow and used contours placed every third image slice, with adjustments on adjacent slices.
- The reference annotation included voxels enclosed by endocardial and epicardial contours while excluding myocardial fat, papillary muscles, and trabeculae carneae.
3. Methods
The method segments the LV myocardium from rest CCTA, encodes its voxel-level characteristics, and classifies patients using clustered features.
- Overall pipeline: The pipeline segments the LV myocardium, extracts characteristics with a CAE, and classifies patients using an SVM.The extracted characteristics are clustered before patient classification.
- Myocardium Segmentation: The two-stage segmentation first localizes the myocardium, then iteratively classifies surface voxels of a rough segmentation.Equidistant voxel classification, Gaussian smoothing, and thresholding produce the rough segmentation before surface refinement.
- Myocardium Segmentation: The multiscale CNN combines orthogonal axial, coronal, and sagittal patches at 49 × 49 and 147 × 147 voxel scales.The smaller patches provide local detail, while the larger downsampled patches provide contextual information.
- Myocardial Characterization: The CAE compresses 48 × 48 voxel axial myocardial patches into encodings and reconstructs the input during unsupervised training.After training, the decoder is removed and the fully connected encoding layer is used as the output representation.
- Patient Classification: The LV myocardium is divided into 500 spatially connected K-means clusters, whose encoding variability characterizes local tissue inhomogeneity.The standard deviation of each encoding across voxels within a cluster is used as the cluster feature.
4. Evaluation
Evaluation assessed segmentation against manual references and assessed patient classification using ROC analysis.
- Myocardium Segmentation: Manual LV myocardium annotation was time-consuming, motivating quantitative evaluation on a subset and qualitative evaluation across all test scans.
- Myocardium Segmentation: Segmentation was quantitatively evaluated with Dice overlap and mean absolute surface distance between automatic and reference LV boundaries.Qualitative segmentation quality was additionally graded by an expert.
- Patient Classification: Patient classification was evaluated using a receiver operating characteristic curve.
5. Experiments and Results
Experiments trained the segmentation CNN on manually annotated scans, tested it independently, and reported strong quantitative overlap with reference myocardium boundaries.
- Myocardium Segmentation: 20 manually annotated scans trained the segmentation CNN, while 20 remaining annotated scans formed an independent quantitative test set.CNN hyperparameters and thresholds were determined using the training set.
- Myocardium Segmentation: Balanced positive and negative voxel patches were used to train the segmentation CNN.The sampling addressed the small myocardial volume and difficult distinction between myocardium and nearby background voxels.
- Myocardium Segmentation: During testing, every fifth voxel was classified in axial, coronal, and sagittal planes before Gaussian smoothing, thresholding, and iterative surface refinement.
- Myocardium Segmentation: 91.4 ± 2.1% Dice coefficient and 0.7 ± 0.1 mm MAD were achieved on the 20-scan segmentation test set.Qualitative evaluation was performed on all 146 test scans.
5.2. Myocardial Characterization
Myocardial characterization used a convolutional autoencoder trained on axial patches and examined reconstructed patches and reconstruction errors in test scans.
- Myocardial Characterization: The CAE was trained and validated on 48 × 48 voxel axial patches extracted around randomly selected myocardial voxels.90% of patches were used for training and 10% for validation.
- Myocardial Characterization: Figure 9 shows randomly selected test-scan input patches alongside their reconstructions and reconstruction errors.
5.3. Patient Classification
Patient classification used clustered encodings of automatically segmented LV myocardium and was evaluated with repeated 10-fold cross-validation. The resulting ROC performance reached an average AUC of 0.74 ± 0.02.
- 126 patients were analyzed after excluding scans used for training or characterization and cases without complete, sufficiently accurate automatic LV myocardium segmentation.The initial cohort contained 166 patients.
- 512 encodings and 500 LV myocardium clusters supplied the characteristics used for SVM-based patient classification.SVM parameters were selected by grid search on the training set only.
- 0.74 ± 0.02 average AUC was achieved across the classification experiments.The ROC curve used a 0.78 FFR cut-off, with shaded variability representing the standard deviation of sensitivity.
5.4. The Effect of Number of Clusters
The number of spatial clusters affected classification performance when the number of encodings was fixed at 512. The highest AUC was obtained with 500 clusters.
- 500 clusters produced the highest AUC among the tested cluster counts.
- The clustering experiment evaluated whether spatial subdivision of the myocardium changed patient-classification performance.
5.5. The Effect of Number of Encodings
The study examined how encoding dimensionality and clustering initialization influenced classification. Performance varied with the number of encodings, while different random seeds produced AUC values from 0.71 to 0.75.
- Number of encodings: 128 and 1024 additional encodings were tested with 500 clusters using separately trained autoencoders.
- Experimental scope: The encoding and clustering experiments assessed sensitivity of patient classification to representation size and clustering choices.
- Seed initialization: 0.71 to 0.75 AUC was obtained across four random seed initializations with 500 clusters and 512 encodings.
5.7. The Effect of FFR Cut-off Values
Classification was evaluated under alternative FFR thresholds because published cut-offs ranged from 0.72 to 0.80. The supplied results also report that correcting segmentation did not significantly change the previously achieved AUC.
- 0.72 to 0.80 was the literature range of FFR cut-off values used to separate functionally significant stenosis from non-significant stenosis.The study used 0.78 as its FFR cut-off.
- 0.72, 0.74, 0.76, 0.80, and 0.85 were additionally tested as FFR cut-off values with 500 clusters and 512 encodings.
- 0.74 ± 0.02 AUC with automatic segmentation was not significantly different from the result after manual segmentation correction.The comparison had p > 0.05.
5.9. Comparison with Other Methods
The proposed myocardium-based method is compared with prior approaches, but differences in patients, scans, and metrics limit direct performance comparisons.
- Different patients, scans, and evaluation metrics mean the reported results provide only an indication of performance differences.
- Table 3 compares evaluated patients and vessels, diagnostic accuracy, and per-patient and per-vessel AUC across previous methods.The methods analyze either coronary blood flow or ischemic changes directly in the LV myocardium.
- Blood-flow methods tend to achieve higher per-patient accuracy than the proposed approach.
- The proposed approach seems to outperform Han et al. (2017) and achieves accuracy equal to the blood-flow method of Min et al. (2012).
6. Discussion and Conclusion
The study presents an automatic LV-myocardium approach for classifying patients with functionally significant stenosis, achieving moderate performance while retaining several methodological and clinical limitations.
- Contribution: The method segments LV myocardium, encodes it with a CAE, and classifies patients using encoding statistics against invasively measured FFR.
- Results: 0.74 ± 0.02 AUC was achieved for identifying functionally significant stenosis using LV-myocardium features alone.Average accuracy, sensitivity, and specificity were 0.71, 0.70, and 0.71, respectively.
- Segmentation: Segmentation was a prerequisite for classification, and limited evaluation showed high segmentation performance, although classification depended on the segmentations.
- Segmentation: Automatic segmentation performed poorly where contrast between the blood pool, LV myocardium, and right ventricle was insufficient.Small segmentation inaccuracies did not significantly influence patient classification in the reported examination.
- Representation: The 2.5D approach limited analysis of volumetric myocardial information because of receptive-field and hardware trade-offs.Future work could evaluate 3D analysis to leverage volumetric information.
- Representation: Fewer CAE encodings caused inferior performance through aggressive compression, while more encodings were non-beneficial because the CAE possibly overfit.
- Clustering: Spatial clustering preserved local ischemic changes, but both global clustering and overly small clusters failed to capture useful regional inhomogeneity.
- Clustering: Random clustering initialization produced average AUC values ranging from 0.71 to 0.75, indicating sensitivity to initialization.