TL;DR
This paper introduces a fully automated pipeline for measuring thoracic aortic aneurysms (TAAs) using a single convolutional neural network (CNN). This approach significantly reduces the time required for these crucial measurements, making them more accessible and potentially improving diagnostic efficiency.
What Problem Does This Paper Address?
Diagnosing and monitoring thoracic aortic aneurysms (TAAs) is critical for preventing potentially fatal ruptures. This process relies heavily on accurately measuring the diameter of the thoracic aorta at specific anatomical locations. Currently, these measurements are performed manually by expert radiologists using specialized 3D imaging software. This manual process is time-consuming, typically taking 15 to 45 minutes per scan, and can be prone to fatigue-induced errors due to its repetitive nature. While previous research has explored automated methods, they often require manual input or only cover a subset of the standard measurement locations. This paper aims to bridge this gap by developing a fully automated system that can measure the aorta at all standard clinical locations with high accuracy and reliability.
Methods
The core of this work is a novel convolutional neural network (CNN) designed for joint segmentation and localization. Instead of separate networks for identifying the aorta's boundaries (segmentation) and pinpointing key anatomical landmarks, this single CNN performs both tasks simultaneously.
Here's a breakdown of the approach:
- 3D UNet Stem: A shared 3D UNet architecture processes the CT scan to extract initial features.
- Dual Heads: This shared stem then branches into two separate "heads":
- Segmentation Head: Predicts a binary mask indicating which voxels belong to the aorta.
- Landmark Localization Head: Predicts the locations of six key anatomical landmarks along the aorta.
- Loss Function: A weighted cross-entropy loss is used for landmark localization, which the authors found to stabilize training and improve accuracy compared to standard mean squared error (MSE) loss. A binary cross-entropy loss is used for segmentation. The total loss is a combination of both.
- Post-Processing: After the CNN inference, the segmentation output is binarized, and the largest connected component is identified as the aorta. Landmark locations are determined by the voxels with the highest probability from the CNN's output.
- Centerline and Diameter Extraction: The segmented aorta and localized landmarks are then used to:
- Generate an aortic centerline using skeletonization and shortest path algorithms (Dijkstra's).
- Localize cross-sections along this centerline at specific points corresponding to the standard clinical measurement locations.
- Estimate the diameter at each cross-section.
- Sinuses of Valsalva (SVS) Correction: A specific refinement is applied to the SVS measurement to mitigate overestimation caused by coronary artery origins. This involves fitting a cubic spline to the boundary, identifying high-curvature points indicative of branches, removing them, and refitting a linear spline to the remaining points.
Datasets
The study utilized a retrospective dataset of 212 CT angiography (CTA) scans from 67 patients diagnosed with TAA and undergoing routine surveillance imaging. These scans were acquired at the University of Michigan.
- Training/Validation: The dataset was split into five disjoint sets for five-fold cross-validation, meaning four sets were used for training and one for testing in each fold.
- Evaluation: Manual measurements for comparison were performed on 60 of the test scans by three independent expert analysts. The average of these manual measurements served as the ground truth.
The paper does not introduce a new benchmark dataset but rather uses a clinically relevant dataset for evaluating their proposed method.
Key Insights and Ideas
- Joint Segmentation and Localization is Key: The paper strongly advocates for a multitask learning approach where segmentation and landmark localization are performed by a single CNN. This joint training yielded statistically better segmentation performance, which is crucial for accurate diameter measurements, even if landmark localization was slightly less accurate than a single-task network.
- Segmentation Accuracy Matters Most: The authors prioritize segmentation accuracy, arguing it's more critical for diameter measurements. The slight trade-off in landmark localization accuracy was deemed acceptable given the overall performance and the importance of accurate segmentation.
- SVS Correction is Necessary: The specific correction applied to the Sinuses of Valsalva (SVS) measurements is important. It demonstrates a practical approach to handle anatomical complexities (like coronary artery origins) that can lead to overestimation in automated measurements.
- Human Variability is a Benchmark: The study highlights that measurement errors are highest in anatomically complex regions like the aortic root and arch. Crucially, they also note that the standard deviation of manual rater measurements is also higher in these same regions, suggesting that the automated system's errors are, to some extent, mirroring inherent human variability and the inherent difficulty of these areas.
Key Results
- Time Savings: The automated pipeline takes approximately 15 seconds per CTA scan on a standard high-performance laptop with a GPU, a dramatic reduction from the 15-38 minutes required for manual measurements.
- Segmentation Performance: The multitask CNN achieved a Dice score of 97.2 ± 2.1% and an ASSD of 0.48 ± 0.65 mm, outperforming previous 3D UNet approaches.
- Diameter Measurement Accuracy:
- The mean absolute error between automated and manual measurements was ≤1 mm at six of nine standard clinical locations.
- Higher errors were observed at the Sinus of Valsalva (SVS), Sinotubular Junction (STJ), and Mid Arch, ranging from 1.43 to 2.18 mm.
- The 95th percentile error was approximately 6 mm, which is better than some prior deep learning pipelines (around 10 mm).
- Landmark Localization: While the multitask network was slightly worse than the single-task network for landmark localization, the absolute differences were small (<1 mm in the 95% confidence interval), and the overall performance was considered good.
Limitations and Future Work
- Anatomical Complexity: The paper acknowledges that measurement accuracy remains suboptimal in regions with highly variable aortic anatomy, such as the SVS and mid-arch. These areas are also challenging for human raters.
- Generalizability: While the pipeline shows promise, its current performance is based on a specific dataset. The authors suggest future work to refine the automated processes to address current inaccuracies and expand its applicability to:
- Non-gated and potentially non-contrast CTA imaging.
- More diverse types of aortic pathology (e.g., abdominal aneurysms, dissections).
- Other imaging modalities like MRI.
- Human Interaction: The authors anticipate that some human interaction will still be required to verify results and correct errors, though the time savings are substantial.
Summary updated