Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Ayman Abaid, Muhammad Ali Farooq, Niamh Hynes, Peter Corcoran, Ihsan Ullah

TL;DR

The study tackles data scarcity in TBAD CTA imaging by using a customized Text-to-Image diffusion pipeline fine-tuned via DreamBooth with LoRA on a limited seed set to generate class-specific TBAD features (TL, FL, FLT, TL+FL) in 2D slices. Outputs are generated at $512×512$ resolution using carefully designed prompts and negative prompts, with 100 training epochs and a small batch size to maximize data diversity while mitigating overfitting, and evaluated with MS-SSIM, FID, Inception-based features, t-SNE, and downstream UNet validation. Results show the synthetic CTAs can exhibit TBAD-related anatomy with fidelity comparable to real data for several classes, though FLT remains challenging to capture in some cases; clinician qualitative assessments corroborate realism alongside occasional noisy samples. Overall, the approach demonstrates a practical pathway to augment and diversify TBAD datasets, addressing privacy concerns and enabling downstream segmentation and prognosis research, with future work extending to 3D CTA generation and video diffusion for prognosis.

Abstract

Stable Diffusion (SD) has gained a lot of attention in recent years in the field of Generative AI thus helping in synthesizing medical imaging data with distinct features. The aim is to contribute to the ongoing effort focused on overcoming the limitations of data scarcity and improving the capabilities of ML algorithms for cardiovascular image processing. Therefore, in this study, the possibility of generating synthetic cardiac CTA images was explored by fine-tuning stable diffusion models based on user defined text prompts, using only limited number of CTA images as input. A comprehensive evaluation of the synthetic data was conducted by incorporating both quantitative analysis and qualitative assessment, where a clinician assessed the quality of the generated data. It has been shown that Cardiac CTA images can be successfully generated using using Text to Image (T2I) stable diffusion model. The results demonstrate that the tuned T2I CTA diffusion model was able to generate images with features that are typically unique to acute type B aortic dissection (TBAD) medical conditions.

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

TL;DR

The study tackles data scarcity in TBAD CTA imaging by using a customized Text-to-Image diffusion pipeline fine-tuned via DreamBooth with LoRA on a limited seed set to generate class-specific TBAD features (TL, FL, FLT, TL+FL) in 2D slices. Outputs are generated at resolution using carefully designed prompts and negative prompts, with 100 training epochs and a small batch size to maximize data diversity while mitigating overfitting, and evaluated with MS-SSIM, FID, Inception-based features, t-SNE, and downstream UNet validation. Results show the synthetic CTAs can exhibit TBAD-related anatomy with fidelity comparable to real data for several classes, though FLT remains challenging to capture in some cases; clinician qualitative assessments corroborate realism alongside occasional noisy samples. Overall, the approach demonstrates a practical pathway to augment and diversify TBAD datasets, addressing privacy concerns and enabling downstream segmentation and prognosis research, with future work extending to 3D CTA generation and video diffusion for prognosis.

Abstract

Stable Diffusion (SD) has gained a lot of attention in recent years in the field of Generative AI thus helping in synthesizing medical imaging data with distinct features. The aim is to contribute to the ongoing effort focused on overcoming the limitations of data scarcity and improving the capabilities of ML algorithms for cardiovascular image processing. Therefore, in this study, the possibility of generating synthetic cardiac CTA images was explored by fine-tuning stable diffusion models based on user defined text prompts, using only limited number of CTA images as input. A comprehensive evaluation of the synthetic data was conducted by incorporating both quantitative analysis and qualitative assessment, where a clinician assessed the quality of the generated data. It has been shown that Cardiac CTA images can be successfully generated using using Text to Image (T2I) stable diffusion model. The results demonstrate that the tuned T2I CTA diffusion model was able to generate images with features that are typically unique to acute type B aortic dissection (TBAD) medical conditions.
Paper Structure (6 sections, 4 figures, 1 table)

This paper contains 6 sections, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Proposed Methodology
  3. Experimental Outcomes
  4. Data Rendering Results using Tuned TBAD T2I Model
  5. Synthetic Data Validation
  6. Conclusion and Future Work

Figures (4)

  • Figure 1: Methodology for training SD models and generating CTA images via diverse image sampling
  • Figure 2: Rendered CT data and advanced data augmentation results with distinct TL, FL, FLT and TL + FL features
  • Figure 3: Real images (first row), synthetic images (second row) and segmentation results predicted by UNet on corresponding synthetic image (third row). In the segmented images, the color green corresponds to TL and red represents FL.
  • Figure 4: t-SNE embedding displaying generated samples (shown in blue) from each class along with their nearest real CTAs (shown in orange).