Source-linked AI summary
Automatic calcium scoring in low-dose chest CT using deep neural networks with dilated convolutions
Nikolas Lessmann, Bram van Ginneken, Majd Zreik, Pim A. de Jong, Bob D. de Vos, Max A. Viergever, Ivana Išgum
TL;DR
This paper addresses the burden of cardiovascular risk assessment in low-dose chest CT screening, where manual calcium scoring is impractical and cardiac motion and image quality complicate analysis. It proposes two consecutive CNNs for concurrent, anatomically labeled detection of coronary, aortic, and valve calcifications, achieving reliable detection and coronary calcium risk categorization across soft and sharp reconstructions.
Problem
Manual calcium scoring alongside lung screening imposes substantial burden, while low-dose image quality and cardiac motion complicate cardiovascular risk assessment.
Method
Two consecutive CNNs use dilated convolutions for a large receptive field and anatomical labeling, followed by false-positive rejection for concurrent calcification detection.
Results
The method achieved good detection performance and high agreement in cardiovascular risk categorization on a large NLST dataset across soft and sharp reconstructions.
Takeaways & Limitations
Automatic scoring can support cardiovascular risk assessment in lung cancer screening and enable evaluation of thoracic aorta and cardiac valve calcium scores in large screening datasets.
Abstract
from arXiv · showhide
Heavy smokers undergoing screening with low-dose chest CT are affected by cardiovascular disease as much as by lung cancer. Low-dose chest CT scans acquired in screening enable quantification of atherosclerotic calcifications and thus enable identification of subjects at increased cardiovascular risk. This paper presents a method for automatic detection of coronary artery, thoracic aorta and cardiac valve calcifications in low-dose chest CT using two consecutive convolutional neural networks. The first network identifies and labels potential calcifications according to their anatomical location and the second network identifies true calcifications among the detected candidates. This method was trained and evaluated on a set of 1744 CT scans from the National Lung Screening Trial. To determine whether any reconstruction or only images reconstructed with soft tissue filters can be used for calcification detection, we evaluated the method on soft and medium/sharp filter reconstructions separately. On soft filter reconstructions, the method achieved F1 scores of 0.89, 0.89, 0.67, and 0.55 for coronary artery, thoracic aorta, aortic valve and mitral valve calcifications, respectively. On sharp filter reconstructions, the F1 scores were 0.84, 0.81, 0.64, and 0.66, respectively. Linearly weighted kappa coefficients for risk category assignment based on per subject coronary artery calcium were 0.91 and 0.90 for soft and sharp filter reconstructions, respectively. These results demonstrate that the presented method enables reliable automatic cardiovascular risk assessment in all low-dose chest CT scans acquired for lung cancer screening.
I. INTRODUCTION
Low-dose chest CT acquired for lung cancer screening can also support cardiovascular risk assessment through calcium scoring, but manual scoring is burdensome and existing methods often depend on segmentation. The paper proposes concurrent vessel-specific detection using dilated-convolution CNNs and a second CNN for false-positive removal.
- Low-dose screening CT can quantify cardiovascular calcifications, but manual scoring adds substantial burden because scans are numerous, noisy, motion-affected, and lack ECG synchronization.
- Automatic methods have focused mainly on dedicated cardiac CT or CTA, with relatively few methods developed specifically for coronary calcium scoring in chest CT.
- The proposed system concurrently detects coronary artery, thoracic aorta, and cardiac valve calcifications in one method to avoid ambiguous results from combining separate systems.
- Voxel-level labeling assigns calcifications to affected vessels, allowing lesions spanning multiple vascular beds to be separated when their prognostic values differ.
- A CNN identifies potential calcifications throughout the image without explicit anatomical segmentation, using dilated convolutions for a large receptive field and a second CNN for candidate refinement.
- The evaluation uses a large, diverse screening dataset containing varied sites, scanner models, and reconstruction algorithms.
II. DATASET
The study uses a diverse subset of low-dose, non-contrast chest CT scans from the National Lung Screening Trial, sampled across common scanner and reconstruction settings and manually annotated for calcifications.
- 1744 scans were selected from 6000 available NLST baseline scans by sampling 25 common combinations of scanner model and reconstruction algorithm.
- The selected scans came from 31 medical centers and 13 scanner models representing four major CT vendors.
- Scans were acquired at breath-hold without contrast enhancement or ECG synchronization, with variable in-plane resolution, slice thickness, and slice spacing.
- Calcifications were manually labeled in every scan by four trained observers and an experienced radiologist, with a 100-scan subset annotated by multiple readers for agreement assessment.
III. METHOD
The method uses two consecutive voxel-classification CNNs: CNN1 detects and anatomically labels candidate calcium using broad context, while CNN2 removes false positives using local image information.
- CNN1 uses a large receptive field to detect calcium and label calcium voxels by anatomical location.
- CNN2 classifies only CNN1-positive voxels as true-positive or false-positive detections using a smaller receptive field and local image information.
A. First stage network (CNN1)
CNN1 analyzes three orthogonal image patches with dilated convolutions to capture broad spatial context while assigning seven anatomical output classes. Auxiliary outputs and probability averaging support classification from the orthogonal subnetworks.
- CNN1 classifies voxels above 130 HU using a purely convolutional architecture that can process entire slices or volumes efficiently.
- Dilated convolutions provide a large receptive field while limiting the number of trainable parameters compared with simply enlarging network depth or kernels.
- The seven CNN1 output classes are LAD, LCX, RCA, TAC, aortic valve, mitral valve, and background.
- CNN1 has a 131 × 131-pixel receptive field, corresponding to roughly a quarter of an axial slice.
- The network processes three orthogonal 2D patches through identical subnetworks, then combines their feature representations to produce posterior probabilities for each voxel.
- Auxiliary softmax outputs enable training from multiple labeled pixels per patch, and their posterior probabilities are combined with the main output by weighted averaging.
B. Second stage network (CNN2)
CNN2 refines CNN1’s candidate detections by separating true calcifications from visually similar false positives, while reconstruction type affects evaluation and scoring.
- B. Second stage network (CNN2): CNN1 labels candidates by anatomical location, including coronary arteries, aorta, and cardiac valves, but artifacts and noise can create false positives.CNN2 is designed to remove these erroneous detections.
- B. Second stage network (CNN2): CNN2 distinguishes true calcifications from false-positive voxels with similar appearance and location.It focuses on local information and finer details rather than broad spatial context.
- B. Second stage network (CNN2): CNN2 uses non-dilated convolutions with max-pooling and analyzes 2.5D inputs within a 65 × 65-pixel receptive field.
- B. Second stage network (CNN2): The method was evaluated separately on soft and sharp filter reconstructions to assess reconstruction-specific calcium detection performance.Sharp reconstructions make edges and noise more prominent than soft reconstructions.
- B. Second stage network (CNN2): Performance assessment quantified calcifications per subject and label using volume and Agatston scores, with F1 scores measuring agreement with manual calcium volumes.Risk categories based on total coronary calcium used linearly weighted κ.
V. EXPERIMENTS AND RESULTS
The experiments used heterogeneous low-dose CT data, standardized image resolution, sequential CNN training, and separate computational stages for scanning and candidate classification.
- V. EXPERIMENTS AND RESULTS: 1687 scans with manual reference standards were split into 1012 training, 169 validation, and 506 testing scans.The scans came from 1459 participants, with all scans from each participant assigned to one subset.
- V. EXPERIMENTS AND RESULTS: 58% of scans used soft filters and 42% used sharp filters.
- V. EXPERIMENTS AND RESULTS: All scans were resampled in-plane to 0.66 mm × 0.66 mm, while slice spacing had already been standardized to 1.5 mm.Predicted label maps were returned to the original resolution using nearest-neighbor interpolation.
- V. EXPERIMENTS AND RESULTS: CNN1 was trained on high-density voxels, whereas CNN2 was trained on candidate voxels classified by CNN1 as any calcification type.Balanced minibatches contained equal numbers of calcium and background voxels.
- V. EXPERIMENTS AND RESULTS: Total computation required 5–7 minutes, with CNN1 averaging 5.5 ± 1 minutes and CNN2 averaging 8.6 ± 8.1 seconds.
A. Detection of CAC
Automatic coronary artery calcium detection performed strongly across soft and sharp reconstructions, and coronary calcium scores supported close agreement in cardiovascular risk categorization.
- A. Detection of CAC: F1 scores for CAC were 0.89 with soft-only training and 0.90 when both soft and sharp reconstructions were used.
- A. Detection of CAC: 0.68 to 0.84: the CAC F1 score increased on sharp reconstructions when sharp images were added to training.
- A. Detection of CAC: Risk-category agreement on soft reconstructions was 90% with soft-only training and 91% with soft-and-sharp training.The categories were based on total CAC Agatston scores.
- A. Detection of CAC: Risk-category agreement on sharp reconstructions increased from 75% to 89% when sharp reconstructions were added to training.
- A. Detection of CAC: Interobserver CAC agreement was high on soft reconstructions, with F1 scores of 0.95 and 0.97 for the second and third observers.For risk categorization, κ was 0.98 and 1.00 in soft reconstructions and 0.92 and 0.99 in sharp reconstructions.
B. Detection of TAC
Thoracic aorta calcification was common in the test scans, and automatic detection achieved similar performance on soft reconstructions while benefiting from sharp-reconstruction training.
- B. Detection of TAC: 450 of 506 scans (89%) contained thoracic aorta calcification according to the observers.
- B. Detection of TAC: The TAC F1 score on soft reconstructions was 0.89 regardless of whether training used soft images alone or both soft and sharp images.
- B. Detection of TAC: 0.71 to 0.81: the TAC F1 score increased on sharp reconstructions when sharp images were added to training.
- B. Detection of TAC: Observers had higher TAC sensitivity than the automatic method but also higher average false positive volume.
C. Detection of cardiac valve calcifications
Automatic detection performed better for coronary artery and thoracic aorta calcifications than for aortic and mitral valve calcifications. Adding sharp reconstructions to training improved valve-calcification F1 scores in sharp-reconstruction images.
- Cardiac valve calcifications were infrequently identified by observers: aortic valve calcifications in 18.2% and mitral valve calcifications in 11.5% of scans.The observer counts were 92 of 506 scans for aortic valve calcifications and 58 of 506 for mitral valve calcifications.
- Adding sharp reconstructions to training increased sharp-image F1 from 0.59 to 0.64 for aortic valve and from 0.52 to 0.66 for mitral valve calcifications.
- The automatic method had lower cardiac-valve detection performance than CAC and TAC detection.Interobserver agreement was particularly low for sharp-reconstruction images, with an F1 score of 0.18 in the reported comparison.
- CNN1 achieved 96.6% sensitivity with 5574 mm3 average false-positive volume on soft reconstructions and 90.7% sensitivity with 18 739 mm3 on sharp reconstructions.These figures were obtained when CNN1 was trained using only images with soft reconstruction kernels.
- CNN2 reclassification reduced sensitivity but substantially lowered average false-positive volume in both soft and sharp reconstructions.With soft-only training, false-positive volume fell by 95.7% and 99.1% in soft and sharp reconstructions, respectively; with mixed training, it fell by 95.4% and 97.0%.
F. Effect of auxiliary output layers and loss terms
The supplied passages describe auxiliary output layers as improving detection performance while slowing learning, alongside comparisons with prior calcium-scoring methods. They also report that existing CAC and TAC methods were evaluated on smaller or less diverse datasets.
- Auxiliary output layers and corresponding loss terms improved CNN1 detection performance but caused the network to learn more slowly.
- The proposed system’s concurrent CAC, TAC, and valve-calcification task had no directly comparable prior method in the supplied literature comparison.The paper states that previous work did not concurrently detect all three categories or detect cardiac valve calcifications.
- Prior CAC methods reported 79% sensitivity at 4 mm3 false-positive volume and 97% sensitivity at 10 mm3 in 231 scans.Those methods were tested on a dataset acquired at one hospital with scanners from a single vendor and one reconstruction setting.
- Prior CAC studies also reported R2 = 0.91 and a Pearson correlation coefficient of 0.86 for agreement with manual scores.These correlation results came from studies using 41 and 1000 scans, respectively.
- For TAC scoring, the paper reported 90% sensitivity, 213 mm3 average false-positive volume, and 89% positive predictive value in 310 soft-reconstruction scans.
H. Voxel-level vs. lesion-level annotation
The paper compares voxel-level and lesion-level calcium annotation and reports that lesion-level annotation can overestimate calcium scores in low-dose CT. Voxel-level labeling was converted to lesion-level labels using region growing and majority voting.
- Voxel-level reference annotations were converted to lesion-level annotations using 3D region growing with the standard 130 HU calcium threshold.Lesions containing different labels were assigned a label by majority voting.
- The annotation examples include correctly detected and labeled calcifications, partially missed voxels, an artificial aortic valve labeled as calcium, and a metal artifact near the aortic wall.
- Lesion-level annotation leads to an overestimation of the calcium score in low-dose CT scans.
VI. DISCUSSION
The proposed two-CNN system concurrently detects and anatomically labels coronary, aortic, and cardiac-valve calcifications in challenging low-dose chest CT. It achieved strong risk categorization and detection performance, while valve and location-disambiguation cases remained more difficult.
- Method: The method uses two consecutive CNNs: CNN1 identifies and spatially labels high-density voxels using dilated convolutions, while CNN2 discards false positives.
- Detection performance: For CAC and TAC detection, the method achieved performance close to interobserver agreement and separated coronary calcifications into LAD, LCX, and RCA.
- Detection performance: LCX, mitral-valve, and aortic-valve calcifications were harder to identify because of anatomical overlap and limited soft-tissue contrast.Observers particularly disagreed on mitral-valve calcifications, while confusion with TAC was common for the aortic valve.
- Error analysis: False positives mainly arose from location mislabeling, low-dose or motion artifacts, and other high-intensity calcifications; CNN2 substantially reduced them while maintaining high sensitivity.
- Annotation and labeling: Voxel classification can partially misclassify lesions but enables splitting calcifications across arterial beds, which can affect cardiovascular risk categorization.
- Method: The method searches the entire image without preprocessing or explicit spatial features, using dilated convolutions to recognize spatial context in three orthogonal 2D patches.
- Generalization and impact: The large, diverse NLST dataset included challenging acquisition conditions, yet the method achieved good detection performance and high risk-categorization agreement across soft and sharp reconstructions.