An Automated Deep Learning Software Tool for Assessing Aortic Remodeling in Aortic Dissection

Technical Summary: An Automated Deep Learning Software Tool for Assessing Aortic Remodeling in Aortic Dissection

Problem Statement
Computed tomography (CT) is the gold standard for evaluating aortic remodeling following surgical interventions for aortic dissection, such as the frozen elephant trunk (FET) procedure. However, large-scale clinical studies required to assess the long-term efficacy of these interventions are currently hindered by the time-consuming, labor-intensive, and costly nature of manual image analysis. While deep learning has advanced 3D volumetric segmentation, clinical decision-making remains reliant on specific 2D metrics (e.g., maximum diameter and cross-sectional area at anatomical landmarks). There is a lack of user-friendly, fully automated software that bridges advanced 3D AI segmentation with the standardized 2D metrics used by clinicians for routine assessment and longitudinal comparison.

Methodology
The authors developed a fully automated, end-to-end software pipeline implemented as an extension module for 3D Slicer (version 5.8.1). The system utilizes deep learning to process CT angiography (CTA) scans without manual intervention.

• Deep Learning Framework: The core segmentation employs the no new (nn)U-Net (version 2) framework, configured for 3D full-resolution analysis. The inference strategy uses an ensemble of five cross-validation folds to maximize robustness.
• Training Datasets: Two open-source datasets were utilized:
  • AortaSeg24: Used to train the "whole-aorta" segmentation model (distinguishing the ascending aorta from the rest of the aorta down to the iliac arteries).
  • ImageTBAD: Used to train the "true lumen" (TL) segmentation model. Due to reported low accuracy in segmenting thrombosed false lumens in the original dataset, the pipeline relies solely on TL predictions; the false lumen area is derived by subtracting the TL area from the total aortic area.
  • TotalSegmentator: Used to automatically identify vertebral bodies (Th8, Th12, L4) for anatomical landmarking.
• Pipeline Architecture: The workflow consists of six sequential steps:
  1. Multi-structure Segmentation: Simultaneous 3D segmentation of the aorta, true lumen, and vertebral bodies.
  2. Reconstruction & Endpoint Definition: The whole aorta is reconstructed via voxel-wise union. The centerline is defined from the ascending aorta centroid to the L4 vertebral level.
  3. Centerline Extraction: A vascular centerline is computed and resampled to 200 control points.
  4. Landmark Projection: Vertebral centroids (Th8 and Th12) are projected onto the aortic centerline based on minimum cranio-caudal distance.
  5. Image Straightening: Curved planar reformation (CPR) is applied to both the CT volume and segmentation masks to align the vessel geometry.
  6. 2D Measurement: Orthogonal 2D cross-sections are extracted at Th8 and Th12. Cross-sectional areas are calculated by counting labeled voxels, and diameters are derived assuming a circular cross-section for the whole aorta.

Study Cohort and Validation
The study was a proof-of-concept retrospective analysis involving 14 patients (28 CT scans: 14 preoperative and 14 postoperative) who underwent FET repair.

• Reference Standard: Two independent observers (a cardiovascular surgeon and a resident) manually measured the aortic and true lumen areas at Th8 and Th12 using a clinical workstation.
• Statistical Analysis: Agreement between the automated pipeline and the manual reference (mean of two observers) was assessed using Intraclass Correlation Coefficients (ICC), Bland-Altman analysis, and Mean Absolute Error (MAE).

Key Results

• Processing Efficiency: The pipeline successfully processed all 28 scans without failure. The mean processing time was approximately 12.55 minutes per scan on a workstation with an NVIDIA Quadro RTX 6000 GPU.
• Segmentation Performance: On five-fold cross-validation, the models achieved high Dice scores for the aorta (0.94 and 0.91) and true lumen (0.95), while thrombosed false lumen segmentation remained low (0.31), validating the decision to exclude it from direct measurement.
• Agreement with Manual Reference: The automated pipeline demonstrated "excellent" agreement (ICC > 0.90) with the manual reference across all metrics:
  • Aortic Area: ICC = 0.913
  • Aortic Diameter: ICC = 0.911
  • True Lumen (TL) Area: ICC = 0.947
  • TL Ratio: ICC = 0.901
• Subgroup Analysis: Excellent agreement was maintained in both preoperative and postoperative phases, as well as at both Th8 and Th12 anatomical levels.
• Remodeling Detection: The pipeline successfully detected the expected pattern of postoperative remodeling, specifically a significant increase in TL area and TL ratio at the Th8 level (adjacent to the distal FET stent), consistent with clinical expectations.

Significance and Claims
The authors position this work as a proof-of-concept study demonstrating the clinical applicability of an automated pipeline for assessing aortic remodeling. The significance of the study lies in three specific contributions:

1. Clinical Scope: Unlike prior studies focused on diagnostic classification or single-timepoint characterization, this pipeline is designed specifically for the longitudinal evaluation of postoperative remodeling following FET repair.
2. Clinically Familiar Output: The system translates complex 3D AI segmentation into the standard 2D metrics (diameter, area, ratio) used by surgeons and radiologists in routine practice, rather than outputting abstract 3D features.
3. Standardized Reproducibility: By using automated vertebral landmark projection, the pipeline ensures measurements are taken at anatomically equivalent positions across timepoints and patients, eliminating a major source of variability in manual longitudinal assessment.

Limitations and Future Work
The authors acknowledge the study's modest scale (n=14) and single-center design as primary limitations. They note that the pipeline's accuracy is constrained by the available training datasets (AortaSeg24 and ImageTBAD), which lack specific FET-related imaging characteristics (e.g., stent metal artifacts). In one case, the model failed to delineate the true lumen within a stented segment due to this "out-of-distribution" anatomy. Future work requires model refinement with larger, curated, multicenter datasets and validation across different scanner platforms to support broader clinical adoption.