Aortic Annulus Detection Based on Deep Learning for Transcatheter Aortic Valve Replacement Using Cardiac Computed Tomography

Yongwon Cho, Soojung Park, Sung Ho Hwang, Min-Seok Ko, Do‐Sun Lim, Cheol Woong Yu, Seong‐Mi Park, Mi‐Na Kim, Yu‐Whan Oh, Guang YangJournal of Korean Medical Science 2023 · DOI

TL;DR

Researchers have developed ADPANet, a specialized 3D deep learning architecture designed to automatically detect the aortic annulus plane in cardiac CT scans. By incorporating a "Permuted AdaIN" layer into a 3D U-Net framework, the model achieves higher accuracy than standard convolutional and transformer-based networks, potentially streamlining the planning process for life-saving heart valve replacements.

What Problem Does This Paper Address?

Transcatheter Aortic Valve Replacement (TAVR) has become a standard treatment for aortic stenosis, but its success depends heavily on precise pre-operative planning. Specifically, clinicians must accurately identify the aortic annulus plane—the "doorway" where the new prosthetic valve will sit.

Currently, identifying this plane is a manual, labor-intensive process. Radiologists must navigate complex 3D CT volumes to find three specific anatomical landmarks: the noncoronary, left coronary, and right coronary cusps (NCC, LCC, and RCC). Because the heart's anatomy varies significantly between patients and the leaflets are incredibly thin (often less than two pixels wide on a CT), automating this task is a significant computer vision challenge.

Methods

The authors introduce ADPANet (Annulus Detection Permuted AdaIN network). The architecture is built upon a customized 3D U-Net, a staple in medical image segmentation, but introduces a novel twist:

  • Permuted AdaIN (pAdaIN): This layer is added to the activation layers of the 3D convolutions. It works by "swapping" or permuting channel-wise statistics (mean and variance) between samples in a batch. This technique helps the model generalize better by preventing it from over-relying on specific image textures or global statistics, essentially forcing it to learn the underlying 3D structure more robustly.
  • Loss Functions: The model was trained using a combination of Dice Loss (for overlap accuracy), Binary Cross-Entropy, and a Root Mean Square Error (RMSE) loss function to measure the distance between the predicted and ground-truth coordinates.
  • Preprocessing: Input data were resampled to a consistent spatial size (256x256x120 voxels) and normalized using z-score intensity normalization.

Datasets

The study utilized a retrospective dataset from a single tertiary medical center:

  • Total Scans: 72 cardiac CT scans from patients who underwent TAVR between 2017 and 2020.
  • Split: The data was divided into Training (51), Tuning (10), and Testing (11) sets.
  • Ground Truth: Three experienced cardiac radiologists manually delineated the annulus planes to create the "gold standard" for training and evaluation.

Key Insights and Ideas

  • Moving Beyond Landmarks: Unlike previous methods that try to find the three specific cusp points first and then calculate the plane, ADPANet is designed to infer the 3D complex structure of the plane directly.
  • The Power of pAdaIN: The inclusion of the Permuted AdaIN layer was the primary driver for performance gains. By shifting channel-wise statistics, the network becomes more resilient to the "noise" of different CT scanners or patient-specific tissue densities.
  • Efficiency in Small Data: Medical datasets are notoriously small. The authors demonstrated that a well-architected CNN (like ADPANet) can actually outperform more complex "Transformer-based" models when data is limited, as Transformers typically require massive datasets to reach their full potential.

Key Results

ADPANet was compared against two benchmarks: a standard "Original" 3D U-Net and a "Transformer-based" 3D U-Net.

  • Accuracy (RMSE): ADPANet achieved the lowest error at 55.078 ± 35.794, compared to 58.862 for the Transformer and 69.656 for the Original U-Net.
  • Overlap (DSC): ADPANet achieved a Dice Similarity Coefficient of 0.496 ± 0.217, outperforming both the Transformer (0.417) and the Original U-Net (0.495).
  • Qualitative Success: The model proved feasible for detecting the complex 3D structure without requiring temporal consistency (i.e., it doesn't need a video/time-series, just a single 3D snapshot).

Limitations and Future Work

While promising, the study notes several hurdles before clinical adoption:

  • Sample Size: The dataset (N=72) is quite small for deep learning. Future work needs to validate these results on much larger, multi-center datasets.
  • Calcification: The researchers specifically excluded patients with heavy aortic valve calcifications to simplify the initial model. However, calcification is extremely common in TAVR patients and can significantly distort image features.
  • Performance Ceiling: A DSC of ~0.50 indicates there is still significant room for improvement in segmentation overlap before the tool can be used without heavy human oversight.

The authors conclude that while ADPANet is a strong step toward automating TAVR planning, the next phase must involve testing the model against the "messier" reality of highly calcified heart valves.

Summary updated

Abstract

Our deep learning framework was feasible to detect the 3D complex structure of the aortic annulus plane using cardiac CT for TAVR. The performance of our algorithms was higher than other convolutional neural networks.
BibTeX
@article{cho2023aortic,
  author = {Yongwon Cho and Soojung Park and Sung Ho Hwang and Min-Seok Ko and Do‐Sun Lim and Cheol Woong Yu and Seong‐Mi Park and Mi‐Na Kim and Yu‐Whan Oh and Guang Yang},
  title = {{Aortic Annulus Detection Based on Deep Learning for Transcatheter Aortic Valve Replacement Using Cardiac Computed Tomography}},
  journal = {Journal of Korean Medical Science},
  year = {2023},
  doi = {10.3346/jkms.2023.38.e306},
}
Figure