TL;DR
This paper introduces a deep learning pipeline for automatically detecting and localizing cervical carotid artery calcifications (CACs) in dental cone-beam computed tomography (CBCT) scans. By combining slice classification with semantic segmentation, the model achieves high accuracy, offering a potential tool to aid dentists in identifying these clinically significant findings.
What Problem Does This Paper Address?
Dental CBCT scans, increasingly used for various dental procedures, can incidentally capture images of the cervical carotid artery. Calcifications in this region (CACs) are a known risk factor for ischemic cardiovascular disease. However, dentists may overlook these findings due to the limited soft tissue contrast in CBCT and the presence of other non-clinically significant calcifications in the area. This study aims to develop an automated system using deep learning to reliably detect and localize these CACs, thereby improving patient care and potentially preventing serious health issues.
Methods
The authors developed a two-step deep learning approach:
- Slice Classification: An Inception V3 convolutional neural network (CNN), pre-trained on ImageNet and then retrained on CBCT data, was used to classify axial slices. The goal was to identify slices located below the C2-C3 intervertebral disc space, effectively narrowing down the region of interest for CAC detection. This classification was refined using linear regression to determine the boundary between slices above and below the C2-C3 disc.
- CAC Detection and Localization: A U-Net architecture, a popular choice for semantic segmentation in medical imaging, was employed for this step. Two methods were explored:
- Method A (Baseline): The U-Net was trained on small, randomly sampled patches (192x192 pixels) from the original axial slices and their corresponding masks. Data augmentation techniques were used to increase the dataset size and prevent overfitting.
- Method B (Improved): This method built upon Method A by using larger, fixed patches (224x224 pixels) from specific regions of interest. Crucially, it incorporated a class-weighted loss function to address the significant class imbalance (many background pixels vs. few calcification pixels), aiming to improve the segmentation of the small calcification regions.
Datasets
The study utilized a dataset of 56 CBCT studies, totaling 15,257 axial slices.
- Training/Validation: 40 CBCT studies (11,252 axial slices) were used for training and validation.
- Testing: 16 CBCT studies (4005 axial slices) were reserved for testing the performance of the trained models.
- Half of the studies in both training and testing sets contained CACs.
- Slices with calcified triticeous cartilage and superior cornu of the thyroid cartilage were excluded to avoid misclassification.
Key Insights and Ideas
- Two-Stage Approach is Effective: Separating the problem into slice classification (locating the relevant anatomical region) and then image segmentation (finding the calcification within that region) proved to be a robust strategy.
- Class Imbalance is a Major Hurdle: The significant disparity between calcification pixels and background pixels in CBCT slices is a critical challenge for segmentation models. Method B's use of a class-weighted loss function directly addressed this, leading to improved performance.
- Patch Sampling Matters: While random patching (Method A) offers generalization, using fixed, larger patches from pre-localized regions (Method B) combined with a better loss function yielded superior results, suggesting a trade-off between generalization and accuracy.
- CBCT for Incidental Findings: The study demonstrates the potential of leveraging dental imaging modalities like CBCT for detecting extra-dental incidental findings with significant clinical implications.
Key Results
- Slice Classification: The Inception V3 classifier achieved results within 15 slices of the radiologist's C2-C3 disc space identification across all test studies.
- CAC Detection (Method A): Achieved a sensitivity of 94.2%, specificity of 96.5%, positive predictive value (PPV) of 56.9%, negative predictive value (NPV) of 99.7%, and diagnostic accuracy of 96.35%.
- Segmentation Performance (Mean IoU):
- Method A (Classifier + U-Net + Multi-patch): 0.7626
- Method B (Classifier + U-Net + 2-patch + class weight balancing loss): 0.8251
- Method B showed an 8.2% improvement in Mean IoU over Method A, highlighting the benefit of the class-weighted loss.
Limitations and Future Work
- Single Source Dataset: The dataset was obtained from a single CBCT machine (iCAT), which may limit the generalizability of the results to scans from different machines or institutions.
- Homogenous Dataset: The dataset was homogenous, and more heterogeneous data from various sources would be beneficial.
- No Power Analysis: A formal power analysis was not conducted to determine the required sample size for model development.
- Limited Sample Size: The study was a pilot study with a limited sample size and no external validation dataset.
- Exclusion of Other Calcifications: The study focused solely on CACs and excluded other neck calcifications, which could enhance the model's clinical utility if incorporated.
Future work suggested includes:
- Incorporating more heterogeneous data from different institutions and machines.
- Utilizing techniques like cross-validation for more robust validation.
- Expanding the dataset to include other types of neck calcifications to improve the model's ability to differentiate and assist clinicians.
Summary updated