1 of 1

Deep Learning for Type-B Aortic Dissection

Ayman Abaid¹, Muhammad Ali Farooq³, Srinivas Ilancheran¹, Talha Iqbal², Niamh Hynes⁴, Peter Corcoran³, and Ihsan Ullah¹^,²

¹School of Computer Science, University of Galway, Ireland, ²Insight SFI Research Centre for Data Analytics, University of Galway, Ireland, 

³School of Engineering, College of Science and Engineering, University of Galway, Ireland, ⁴School of Medicine, at University of Galway, Galway, Ireland

​

Background & Motivation

Type B aortic dissection (TBAD) is a serious cardiovascular condition caused by a tear in the descending aorta. Accurate lumen segmentation is essential for effective diagnosis and prognosis. Computed Tomography Angiography (CTA) is widely used for detailed visualization of the aorta and surrounding structures, but manual segmentation is time-consuming and costly, underscoring the need for more efficient diagnostic methods. While deep learning has significantly advanced medical image segmentation, developing models for new datasets remains challenging due to the need for precise annotations. Our research focuses on enhancing TBAD diagnosis and prognosis through advanced deep learning methods. Below is a summary of our work so far.

To train, or not to train: exploring foundation models for TBAD segmentation

Exploratory analysis of Type B Aortic Dissection (TBAD) segmentation in 2D CTA images using various kernels

Synthesizing CTA Image Data for Type-B Aortic Dissection using Stable Diffusion Models

1

2

3

Acknowledgements

These publications have emanated from research conducted with the financial support of Science Foundation Ireland under Grant Number [SFI/12/RC/2289_P2] through the Insight SFI Research Centre for Data Analytics. Additionally, these publications have emanated from research conducted with the financial support of Taighde Éireann – Research Ireland under Grant No. 18/CRT/6223.

References

1. Abaid, A., Ilancheran, S., Iqbal, T., Hynes, N., & Ullah, I. (2024). Exploratory analysis of Type B Aortic Dissection (TBAD) segmentation in 2D CTA images using various kernels. Computerized Medical Imaging and Graphics, 118, 102460.
2. Abaid, A., & Ullah, I. (2024, August). To train, or not to train: exploring foundation models for TBAD segmentation. In IET Conference Proceedings CP887 (Vol. 2024, No. 10, pp. 323-326). Stevenage, UK: The Institution of Engineering and Technology.
3. Abaid, A., Farooq, M. A., Hynes, N., Corcoran, P., & Ullah, I. (2024, July). Synthesizing cta image data for type-b aortic dissection using stable diffusion models. In 2024 46th Annual International Conference of the IEEE Engineering in Medicine and Biology Society (EMBC) (pp. 1-5). IEEE.

In this work, we explore 2D Convolutional Neural Networks (CNNs) for slice-by-slice segmentation of true lumen, false lumen, and false lumen thrombus in CTA images. We evaluate three 2D UNet variants—baseline UNet, UNet with atrous convolutions, and a custom UNet with position-oriented weighting—against a state-of-the-art 3D UNet. Our UNet with a VGG19 encoder performed best, achieving a mean Dice score of 80.48% and an IoU of 72.93% (Table 1). While our 2D models excel in lumen segmentation, they show lower accuracy in thrombus segmentation compared to the 3D UNet.

Table1: Comparison of segmentation performance in terms of Dice score (%): Proposed vs. State-of-the-Art Models

Recent advances in computer vision have been fueled by vision foundation models (FMs). Models like the Segment Anything Model (SAM) are now being applied to medical imaging. This study evaluates three segmentation FMs for aortic segmentation on CTA images, comparing their performance to traditional methods. While these models show promise in improving efficiency, our findings suggest that they still fall short of outperforming conventional segmentation models out of the box.

Table 2: Performance comparison of UniverSeg, MedSAM, and U-Net on CTA image segmentation based on the mean Intersection over Union (mIoU) score and mean Dice coefficient score (mDSC). ‘s' denotes the support size.

Table 3: Performance comparison of the classifier using real data versus a hybrid dataset in terms of overall accuracy and class-wise accuracy.

Table 4: Performance comparison of the u-net model with and without synthetic data pretraining, reporting metrics such as mean intersection over union (IoU), and class-wise IoU

This study proposes a data generation pipeline, DiffusionTBAD, which generates synthetic CTA images by fine-tuning Text-to-Image (T2I) stable diffusion (SD) models. Given the small size of medical datasets, we employ the few-shot learning method DreamBooth, combined with LoRA, to fine-tune the T2I SD model. We generate CTA images across five classes, each representing the presence or absence of the lumen, with each image accompanied by a short textual prompt. A comprehensive evaluation, including quantitative analysis and clinician assessment, confirms that T2I SD models can generate realistic CTA images. Additionally, we use the generated data to augment real datasets for downstream tasks such as classification and segmentation. Results in Tables 3 and 4 show that the proposed methodology effectively captures key features of TBAD, demonstrating its potential to address data scarcity in cardiovascular imaging.

​

.

Figure 1:  Methodology for training SD models and generating CTA images via diverse image sampling

Google Scholar Link