TLDR: AortaDiff is a new AI framework that uses diffusion models to simultaneously create synthetic contrast-enhanced CT images and segment abdominal aortic aneurysms from non-contrast CT scans. This unified approach improves both image quality and segmentation accuracy, especially for the challenging thrombus, and works effectively even with limited labeled data. It reduces the need for risky contrast agents, making AAA assessment safer and more sustainable.
Assessing abdominal aortic aneurysms (AAA) traditionally relies on contrast-enhanced CT (CECT) scans. However, the iodinated contrast agents used in CECT pose several risks, including potential kidney damage, patient allergies, and environmental concerns. To address these issues, researchers are exploring ways to perform AAA assessment using only non-contrast CT (NCCT) scans.
Existing deep learning methods often tackle this problem in multiple stages: first generating a synthetic CECT image, and then performing segmentation on that image. This multi-stage approach can lead to errors accumulating along the pipeline and doesn’t fully leverage the shared anatomical information between image generation and segmentation tasks. Other single-task methods can produce inconsistent results, and earlier multi-task models based on Generative Adversarial Networks (GANs) sometimes suffer from overly smooth outputs and training instability.
A new framework called AortaDiff aims to overcome these limitations by offering a unified, multi-task deep learning solution. Developed by a team from the University of Oxford, Technical University of Munich, ELLIS Institute Finland, and Aalto University, AortaDiff simultaneously generates synthetic CECT images from NCCT scans and segments the aortic lumen and thrombus. This integrated approach ensures that both tasks benefit from shared learning, leading to more accurate and consistent results.
AortaDiff integrates conditional diffusion models (CDM) with multi-task learning, allowing for end-to-end optimization of both image synthesis and anatomical segmentation. Unlike some previous multi-task diffusion models, AortaDiff doesn’t require any initial predictions, such as a coarse segmentation mask. It also shares both encoder and decoder parameters across its tasks, promoting a rich, shared understanding of the anatomy.
A key innovation is its semi-supervised training strategy. This is particularly valuable in clinical settings where obtaining fully labeled data (NCCT, CECT, and segmentation masks) is challenging and expensive. AortaDiff can learn effectively from datasets where many scans only have image pairs but lack detailed segmentation labels, leveraging the entire dataset to improve its performance.
The framework also addresses practical challenges like spatial misalignments between NCCT and CECT scans, which can occur due to patient movement. AortaDiff employs an aorta-focused generation strategy with a background preservation technique. This means the model concentrates its generative efforts specifically on the aortic region, preventing it from learning spurious translations in surrounding tissues and ensuring stable performance.
In experiments conducted on the Oxford Abdominal Aortic Aneurysm (OxAAA) dataset, AortaDiff demonstrated significant improvements. For image synthesis, it achieved a PSNR of 25.61 dB, outperforming a single-task CDM. More importantly, for anatomical segmentation, AortaDiff improved the lumen Dice score to 0.89 (from 0.87 for nnU-Net) and the challenging thrombus Dice score to 0.53 (from 0.48 for nnU-Net). These enhancements in segmentation led to more accurate clinical measurements, reducing the lumen diameter Mean Absolute Error (MAE) to 4.19 mm (from 5.78 mm) and the thrombus area error to 33.85% (from 41.45%).
Also Read:
- Precise Image Modifications: A New Approach for Counterfactual Synthesis in Medical Imaging
- U-DFA: A New Approach to Medical Image Segmentation with Dual Fusion Attention
The semi-supervised variant of AortaDiff, AortaDiff-P, showed the best performance across both tasks, highlighting the effectiveness of its training strategy in scenarios with limited labels. This research represents a significant step towards creating a “digital contrast” for CT, reducing the reliance on iodinated contrast agents. Such AI-driven workflows can enhance patient safety, lower procedural costs, and minimize the environmental impact of medical waste, contributing to a more sustainable healthcare system. You can find more details about this research paper here: AORTADIFF: A Unified Multitask Diffusion Framework for Contrast-Free AAA Imaging.
