Volumetric Analysis of Uncomplicated Type B Aortic Dissection Using an Automated Deep Learning Aortic Zone Segmentation Model
Jonathan R Krebs, Muhammad Imran, Brian Fazzone, Chelsea Viscardi, Benjamin Berwick, Griffin Stinson, Evans Heithaus, Gilbert R Upchurch, Jr., Wei Shao, Michol A Cooper
University of Florida, Gainesville, FL
Background:Despite 20 years of international clinical trials and evaluation, there are no clear guidelines for the optimal treatment algorithm for patients presenting with acute uncomplicated type B aortic dissection (auTBAD). A significant contributing factor is the lack of image analysis techniques that accurately predict aortic growth. Aortic diameter >40mm is the only prospectively validated high-risk anatomic feature predicting need for future intervention. However, diameter measurements are subject to high inter-reader variability and do not capture the three-dimensional (3D) nature of aortic growth. Volumetric CT angiography has been suggested as a means of overcoming these limitations. Deep learning techniques have shown excellent performance in 3D medical image analysis, but have not been applied to TBAD nor incorporated SVS-defined aortic zones. The purpose of this study was to establish a trained, automatic deep learning aortic zone segmentation model to facilitate performance of an aortic zone volumetric comparison between medically managed auTBAD patients with and without aortic growth. We hypothesized that differences in baseline volumetrics, particularly in zones of the thoracic aorta, would be present in patients that experienced aortic growth compared to those that did not. Methods:Patient identification: Using a prospectively maintained institutional database, a retrospective review of patients admitted to our center with a diagnosis of code of aortic dissection (ICD-9 codes 441.01, 441.03; ICD-10 codes I71.00, I71.01) between 10/2011 and 3/2020 was performed. Type A dissection, intramural hematoma, penetrating aortic ulcer, and chronic aortic dissection patients were excluded. Uncomplicated TBAD patients were identified based on the absence of malperfusion, rupture, rapid degeneration, or refractory pain. All uncomplicated TBAD patients were medically managed without surgical intervention at index admission. Imaging was then reviewed and patients without high-resolution surveillance imaging beyond three months from discharge after index hospitalization were excluded. To optimize our 3D training model, patients with bovine arch or aberrant arch anatomy were also excluded. Data collection: Patient demographics, comorbidities, and hospital course were obtained from the electronic medical record. A board-eligible radiologist and vascular surgeon analyzed imaging characteristics from two CT scans: the baseline CTA at index admission, and either the most recent surveillance CTA, or the most recent CTA prior to an aortic intervention if one was performed. Patients were then stratified into two groups: aortic growth (AG) (≥5mm/year) and no aortic growth (NAG) (<5mm/year). Manual Segmentation and Segmentation Model: Baseline and surveillance CTAs were deidentified and downloaded from our institutional software platform to allow for importation to 3D Slicer, a free open-source software package for medical image analysis. CT scans were randomly partitioned into training (80%), validation (10%), and testing (10%) for unbiased evaluation. After importing the deidentified images into 3D Slicer, the 11 aortic zones were manually segmented based on SVS/STS criteria. A Gaussian smoothing filter was applied to reduce jaggedness and enhance 3D continuity. A SkipXNet architecture was then used to generate the segmentation output and optimized with the Adam optimizer and a combined cross-entropy and soft dice loss function. Using the custom optimized-segmentation output, the volume of each aortic zone was computed using the segment statistics function within 3D Slicer. Statistical Analysis: Primary comparisons were made of the volumes in the different aortic zones between Aortic Growth (AG) and No Aortic Growth (NAG) groups using R statistical package. Mean comparisons were made using t-test/Man-Whitney test and Chi-square/Fisher Exact test as appropriate with a p-value of ≤0.05 considered significant.
Results: Of the 159 patients treated for uncomplicated TBAD, 76 patients had the requisite surveillance imaging for inclusion. An additional 17 patients were excluded due to aberrant aortic arch anatomy. Of 59 included patients, aortic growth was observed in 33 (56%) patients. Thoracic growth rate in the AG group was 14.8±13.5 mm/year compared to 0.66±5.4 mm/year in the NAG group (p<0.01). Abdominal growth in the AG group was 4.9±5.0, compared to 1.2±2.2mm/year in the NAG group (p<0.01). Median duration between baseline and interval CT was 1.07 years (IQR 0.38-2.57) and longer in the NAG group at 2.2 years (IQR 1.2-4.0) versus 0.65 years (IQR 0.26-1.09) in the AG group (p<0.01). Post-discharge surgical intervention was performed in 22% (n=13) of patients at a mean of 1.5±1.2years, with no difference between AG (21%) and NAG (23%) groups (p=1.0). There were no differences between AG and NAG groups with respect to age, BMI, sex, race, baseline comorbidities, admission mean arterial pressure, and number of discharge antihypertensive medications (p>0.05 for all). Baseline CT characteristics: Mean baseline maximum thoracic aortic diameter was greater in the NAG group (44.3±8.5) than the AG group (40.3±5.0) (p=0.03). Mean baseline abdominal aortic diameter was also greater in the NAG group (34.3±5.7) than the AG group (31.7±4.1) (p=0.05). There were no differences in false lumen patency, incidence of false lumen diameter >22mm, or incidence of lesser curve entry tear between AG and NAG groups (p>0.05). Interval CT characteristics: At interval CT, mean thoracic aortic diameter was similar in the AG group (48.8±8.8) compared to the NAG group (48.8±10.0) (p=1.0). There was also no difference in abdominal aortic diameter between the AG group (36.7±6.7) and the NAG group (38.7±7.6) (p=0.27). There remained no differences in false lumen patency or false lumen diameter >22mm between AG and NAG groups (p>0.05). Segmentation model: The performance of the 3D-model is measured based on the number of overlapping pixels between the physician-annotated ground-truth (actual) region and the model predicted region. Performance is scored based on the Dice coefficient which ranges from 0 to 1 with 1 indicating a perfect match. Dice coefficient was tested using random sample of the training dataset with an overall performance of 0.73. Performance was best in Zone 4 (0.82), Zone 5 (0.88), and Zone 9 (0.91), with worse performance in Zone 3 (0.54), Zone 6 (0.60), Zone 7 (0.65), and Zone 8 (0.70). Aortic Zone Volumetrics: Mean aortic zone volumes from baseline CT are shown in Table 1. There were no differences in baseline zone volumes between the AG and NAG groups (p>0.05 for all). Conclusions:This is the first description of an automatic deep learning aortic segmentation model incorporating SVS-defined aortic zones and trained using a real-world dataset of patients with uncomplicated Type B aortic dissection. The open-source, trained model demonstrates high concordance to ground-truth aorta with particularly strong performance in zone 4, zone 5, and zone 9. With exposure to additional imaging and anatomic variation of the visceral and renal vessels, the performance in zones 6 through 8 is expected to improve.
In our sample, there did not appear to be differences in baseline aortic zone volumes between patients with and without aortic enlargement over time. It is notable that we also found no differences between groups when looking at other high-risk features like maximum aortic diameter >40mm, false lumen >22mm, false lumen thrombosis status, and presence of lesser curve entry tear. Baseline maximum aortic diameter, for example, was actually larger in the group without aortic growth over time. This speaks to the limitations of our patient sample size and underscores the need to expand the use of our model to better identify if baseline aortic zone volume thresholds are associated with growth over time. This can be performed in several ways: the most obvious is with a larger sample of uTBAD patients, although a deidentified set of TBAD images can be difficult to obtain and is not publicly available at this time. In other cases, predictive models can use “in silico” data, or large volumes of simulated data generated from computer models to pre-train the model prior to patient image exposure to offset the limitations of a smaller dataset.Beyond the limitations of our sample size, there are several others to consider. We excluded patients with aberrant anatomy to optimize model performance but further efforts will need to incorporate aberrant anatomy. The sample had a considerable number of patients with an only three-month interval between baseline and surveillance CT. As a result, our sample may have been biased towards patients that experienced rapid aortic growth which has previously been shown to display distinct phenotypic features. Additionally, there is not a standard definition of aortic growth in the literature which limits our ability to compare growth patterns and characteristics between similar work from prior centers. While aortic volume may better reflect the 3-D nature of aortic growth compared to diameter alone, our model did not consider dissection flap morphology or false lumen volume. The next model iteration will add these layers to the existing zone segmentation to better understand their contributory role, if any, towards aortic growth over time.
Table 1: Patient demographics, baseline CT reference characteristics, and model calculated aortic zone volumes (mm³) represented as mean±sd, median (IQR), or n (%).
All patients(n=59) | Aortic Growth(n=33) | No Aortic Growth(n=26) | p-value | |
Index admission age (years) | 59.2±14.6 | 57.7±15.4 | 61.2±13.7 | 0.35 |
Female Sex | 20 (34) | 12 (36) | 8 (31) | 0.78 |
Body Mass Index | 30.7±6.7 | 30.9±5.1 | 30.5±8.3 | 0.81 |
Mean arterial pressure on initial presentation | 95.5±28.4 | 101±30 | 89±25 | 0.10 |
Number of discharge antihypertensive medications | 3 (1.1) | 3.0 (1.1) | 3.0 (1.3) | 0.83 |
Time from discharge to most recent CTA or CTA prior to TEVAR (years) | 1.07 (IQR 0.38-2.57) | 0.65 (IQR 0.26-1.09) | 2.2 (IQR 1.2-4.0) | <0.01 |
TEVAR during surveillance period | 13 (22) | 7 (21) | 6 (23) | 1.0 |
Growth rate (mm/year) | 8.6±12.8 | 14.8±13.5 | 0.66±5.4 | <0.01 |
Baseline Maximum Thoracic Aortic Diameter (mm) | 42.0±7.0 | 40.3±5.0 | 44.3±8.5 | 0.03 |
Baseline Maximum Abdominal Aortic Diameter (mm) | 32.9±5.0 | 31.7±4.1 | 34.3±5.7 | 0.05 |
Thrombosed False Lumen | 5 (8) | 1 (3) | 4 (15) | 0.13 |
Partially Thrombosed False Lumen | 27 (46) | 18 (55) | 9 (35) | |
Patent False Lumen | 27 (46) | 14 (43) | 13 (50) | |
Maximum diameter >40mm | 36 (61) | 17 (52) | 19 (73) | 0.11 |
Maximum false lumen diameter >22mm | 33 (56) | 16 (48) | 17 (65) | 0.60 |
Lesser curve entry tear | 3 (5) | 2 (6) | 1 (4) | 1.0 |
Zone 3 | 19,116±7,038 | 20,304±7,230 | 17,974±6,790 | 0.24 |
Zone 4 | 84,383±39,311 | 89,510±43,116 | 79,454±35,420 | 0.37 |
Zone 5 | 148,985±60,201 | 147,121±66,855 | 150,776±54,321 | 0.83 |
Zone 6 | 11,505±4,896 | 10,604±3,961 | 12,371±5,594 | 0.20 |
Zone 7 | 8,054±4,554 | 8,734±5,431 | 7,400±3,502 | 0.30 |
Zone 8 | 8,317±5,189 | 8,966±6,734 | 7,693±3,083 | 0.39 |
Zone 9 | 40,938±18,970 | 38,818±15,635 | 42,977±21,822 | 0.43 |
Back to 2024 Abstracts