AI-Driven Assessment of Aortic Dimensions Based on Contrast-Enhanced MRA of the Thoracic Aorta (RF_TH_164)

Presenting Author(s)

• CA

  Charilaos Apostolidis, MEng

  PhD Candidate
  Northwestern University

Primary Author(s)

• CA

  Charilaos Apostolidis, MEng

  PhD Candidate
  Northwestern University

Co-Author(s)

• HB

  Haben Berhane, BSc

  Graduate student
  Northwestern University
• DD

  David Dushfunian, MD

  Clinical Research Associate
  Northwestern University Feinberg School of Medicine
• 

  Ethan M. I Johnson, PhD

  Clinical Research Associate
  Northwestern University
• CP

  Courtney Pfister, BSc

  Student
  Northwestern University
• 

  Bradley D. Allen, MD, MSc, FSCMR

  Assistant Professor, Cardiovascular and Thoracic Imaging
  Northwestern University
• AK

  Aggelos Katsaggelos, PhD

  Professor
  Northwestern University
• 

  Michael Markl, PhD

  Professor
  Northwestern University Feinberg School of Medicine

Figure 1: Overview of the analysis workflow including DL-based automated segmentation and aortic dimension quantification method. There are four main tasks: (a) segmentation, (b) plane placement, (c) diameter calculation, (d) volume calculation. A CEMRA scan is input to a trained 3D Dense U-Net which produces a segmentation mask of the aorta. The 3D U-Net downsamples the input images through a series of 3D convolutions and pooling operations. This allows it to capture global information and the overall structure of the images. It then upsamples these features back to the original size using transposed convolutions, capturing local information. Skip connections propagate the higher-level features derived at earlier stages, allowing the network to process them together with the lower-level ones, which allows it to effectively capture the anatomy. The centerline of the segmentation mask is automatically calculated, and planes are manually placed orthogonally, according to the guideline[3]-indicated locations. Major- and minor-axis diameters at the locations of the planes, as well as volume are automatically calculated.

Figure 2: Comparison of the best, median and worst automatic segmentation masks, as ranked by their DSC. Manual plane placement is also demonstrated (except for the worst-DSC mask, which was excluded from diameter and volume analysis).

Figure 3: Comparison of anatomical metrics (ascending aorta diameter, aortic arch diameter, descending aorta diameter, volume) for DL based vs. manual analysis in the final testing cohort (N = 42). Diameter values are of the major axis. Median values, IQR for each metric, the median percentage difference between corresponding manual and AI values, and the results of the Wilcoxon Signed-Rank (WSR) test (test statistic and p-value) are presented. Bland-Altman plots for each metric are included.


Background: Aortic valve disease (AVD) affects 2-3% of the US population, leading to severe complications (e.g. aneurysm, dissection) [4]. To guide patient management, contrast-enhanced MR angiography (CEMRA) is often used to monitor aortic dimensions and growth rates [3]. However, CEMRA-based assessment of aortic dimensions relies on manual measurements at guideline[3]-defined anatomic landmarks, which are time-consuming with variable reproducibility. There is an unmet clinical need for robust, automated aortic anatomy quantification from CEMRA scans. In this study, we report on the feasibility of a deep learning (DL) method for fully automated, 3D aortic segmentation, followed by highly automated quantification of aortic dimensions at multiple anatomic locations. We hypothesize that this AI-driven analysis can accurately reproduce manual quantification of aortic volumes and dimensions.


Methods: We manually segmented 172 CEMRA scans of our bicuspid aortic valve (BAV) patient cohort (52F, 49.3±16.2 yrs) and used 128 of them for model training and 44 for validation. The scans had a flip angle of 25-40, TE of 0.89-1.23ms and spatial resolution of (0.63-1.3)x(0.63-1.3)x(1-2)mm3. Manual segmentations were performed using Materialise Mimics.

 

This work expands our prior efforts of using a 3D U-Net [2] with densely connected layers for segmentation. The 3D U-Net consists of an encoder-decoder structure with dense connections (Fig. 1). Manual segmentations of CEMRA scans were used as ground truth. The ground truth was used to calculate the loss (composite cross-entropy and Dice loss), which guides the model in minimizing the difference between predicted and ground-truth segmentations. Segmentation success was evaluated using the Dice similarity coefficient (DSC).

We calculated aortic diameters (the major axis of the cross-sectional ellipse) by placing cross-sectional planes along the centerline of the validation set masks. Measurement locations were mid-ascending aorta (AAo), arch, mid-descending aorta (DAo). Cases where the segmentation DSC was under 0.70 were excluded. We compared diameters and total aortic volume between AI-predicted and ground truth (manual) segmentations. A Wilcoxon Signed-Rank test was performed to test the hypothesis.


Results: Average AI segmentation time was 1.6±0.2s. AI-predicted 3D segmentation masks were in close agreement with manual ground truth (median DSC=0.94, IQR = [0.90-0.96]. The best, median and worst AI segmentations (ranked by DSC) are presented in Fig. 2. The worst AI segmentation (DSC=0.43) was an outlier.


As shown in Fig. 3, DL-based quantification of aortic dimensions showed excellent agreement with human observers with minimal differences in diameters (<4%) and volumes (<6%).  No significant difference between manual- and AI-derived metrics was found (p-value of 0.35, 0.13, 0.77, 0.12, for AAo, arch, DAo, and volume).


Conclusion: CEMRA-based quantifications of aortic diameters and volume are consistent between DL-generated aortic segmentations and manually-created segmentations. This suggests that this highly automated, AI-driven approach could potentially replace cumbersome manual analysis, substantially decrease processing times and improve reproducibility. Increasing the training set size and heterogeneity, as well as methodological refinements will improve the DL model’s robustness to noise artifacts and poorer contrast. Automation of anatomical landmark localization and plane placement will also be pursued.

Figure 1: Overview of the analysis workflow including DL-based automated segmentation and aortic dimension quantification method. There are four main tasks: (a) segmentation, (b) plane placement, (c) diameter calculation, (d) volume calculation. A CEMRA scan is input to a trained 3D Dense U-Net which produces a segmentation mask of the aorta. The 3D U-Net downsamples the input images through a series of 3D convolutions and pooling operations. This allows it to capture global information and the overall structure of the images. It then upsamples these features back to the original size using transposed convolutions, capturing local information. Skip connections propagate the higher-level features derived at earlier stages, allowing the network to process them together with the lower-level ones, which allows it to effectively capture the anatomy. The centerline of the segmentation mask is automatically calculated, and planes are manually placed orthogonally, according to the guideline[3]-indicated locations. Major- and minor-axis diameters at the locations of the planes, as well as volume are automatically calculated.

Figure 2: Comparison of the best, median and worst automatic segmentation masks, as ranked by their DSC. Manual plane placement is also demonstrated (except for the worst-DSC mask, which was excluded from diameter and volume analysis).

Figure 3: Comparison of anatomical metrics (ascending aorta diameter, aortic arch diameter, descending aorta diameter, volume) for DL based vs. manual analysis in the final testing cohort (N = 42). Diameter values are of the major axis. Median values, IQR for each metric, the median percentage difference between corresponding manual and AI values, and the results of the Wilcoxon Signed-Rank (WSR) test (test statistic and p-value) are presented. Bland-Altman plots for each metric are included.