Region of interest detection for efficient aortic segmentation

L Giordano, Ine Dirks, Tom Lenaerts, Jef Vandemeulebroucke

Abstract

Thoracic aortic dissection and aneurysms are the most lethal diseases of the aorta. The major hindrance to treatment lies in the accurate analysis of the medical images. More particularly, aortic segmentation of the 3D image is often tedious and difficult. Deep-learning-based segmentation models are an ideal solution, but their inability to deliver usable outputs in difficult cases and their computational cost cause their clinical adoption to stay limited. This study presents an innovative approach for efficient aortic segmentation using targeted region of interest (ROI) detection. In contrast to classical detection models, we propose a simple and efficient detection model that can be widely applied to detect a single ROI. Our detection model is trained as a multi-task model, using an encoder-decoder architecture for segmentation and a fully connected network attached to the bottleneck for detection. We compare the performance of a one-step segmentation model applied to a complete image, nnU-Net and our cascade model composed of a detection and a segmentation step. We achieve a mean Dice similarity coefficient of 0.944 with over 0.9 for all cases using a third of the computing power. This simple solution achieves state-of-the-art performance while being compact and robust, making it an ideal solution for clinical applications.

Introduction

The aorta is the body’s primary "highway" for blood, and conditions like dissections (tears) or aneurysms (bulges) are life-threatening. Doctors need precise 3D segmentations of the aorta from CT scans to plan treatments, but doing this manually is exhausting and prone to bias.

While deep learning models can automate this, they face a significant challenge with 3D medical images: full images are often too large for standard GPU memory, while the standard solution of using small patches loses the "big picture" context. This loss of context leads to errors in complex geometries where the aorta is distorted. Furthermore, high-end models like nnU-Net are computationally expensive, making them difficult to deploy in standard clinical settings.

Methods

The authors propose a two-stage Cascade Model to address these challenges.

Cascade Architecture

The first stage is a Detection Stage that uses a simplified multi-task model with a SwinUNETR backbone. At the network's bottleneck, a simple Fully Connected Network (FCN) predicts a single bounding box. This model is trained to simultaneously segment the aorta and predict its location. Once the ROI is detected, the image is cropped, and a second, more focused SwinUNETR model performs high-quality segmentation on the smaller volume.

Overview of the cascade model architecture, featuring the multi-task detection stage and the secondary focused segmentation stage.

Datasets and Training

The study combined two major open-access datasets: the AVT (Aortic Vessel Tree) dataset (56 patients, primarily healthy) and ImageTBAD (100 patients with Type B aortic dissections). All images were resampled to a 2 mm isotropic resolution and standardized to a 256x256x256 voxel size.

Key Insights and Results

The Mass Heuristic

A key innovation of this study is the "mass" heuristic. The authors observed that the number of aortic voxels in a horizontal slice drops sharply right above the aortic arch. This anatomical landmark was used to automatically generate ground-truth bounding boxes for training, removing the need for manual labeling.

Illustration of the mass heuristic used for ROI detection, showing the normalized mass distribution along the axial direction and the corresponding 3D bounding box.

Performance and Efficiency

By cropping to an ROI, the model gains significant anatomical context, achieving 42.2% patch coverage compared to only 12.5% for nnU-Net. This extra context is why the model performs better on difficult cases where the aorta's shape is highly distorted.

  • Accuracy: The model achieved a mean Dice Similarity Coefficient of 0.944, with every test case scoring above 0.9.
  • Precision: The Hausdorff Distance (HD95) was nearly halved compared to nnU-Net (2.6 mm vs 4.3 mm).
  • Computational Savings: The model required only 10.3 GB of GPU memory (vs. 36.4 GB for nnU-Net) and completed training in 18.5 hours (vs. 27 hours).

Conclusion

This cascade approach provides a robust and efficient solution for clinical aortic analysis. By simplifying the detection task to a single ROI, the researchers created a lean tool that outperforms standard industry methods while using significantly fewer resources. Future work may extend this "detect-then-segment" pipeline to other single organs like the liver or heart.

Summary updated

Abstract

Thoracic aortic dissection and aneurysms are the most lethal diseases of the aorta. The major hindrance to treatment lies in the accurate analysis of the medical images. More particularly, aortic segmentation of the 3D image is often tedious and difficult. Deep-learning-based segmentation models are an ideal solution, but their inability to deliver usable outputs in difficult cases and their computational cost cause their clinical adoption to stay limited. This study presents an innovative approach for efficient aortic segmentation using targeted region of interest (ROI) detection. In contrast to classical detection models, we propose a simple and efficient detection model that can be widely applied to detect a single ROI. Our detection model is trained as a multi-task model, using an encoder-decoder architecture for segmentation and a fully connected network attached to the bottleneck for detection. We compare the performance of a one-step segmentation model applied to a complete image, nnU-Net and our cascade model composed of a detection and a segmentation step. We achieve a mean Dice similarity coefficient of 0.944 with over 0.9 for all cases using a third of the computing power. This simple solution achieves state-of-the-art performance while being compact and robust, making it an ideal solution for clinical applications.
BibTeX
@misc{giordano2025region,
  author = {L Giordano and Ine Dirks and Tom Lenaerts and Jef Vandemeulebroucke},
  title = {{Region of interest detection for efficient aortic segmentation}},
  year = {2025},
  eprint = {2601.08683},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.08683},
}
Figure