Daily Arxiv

This is a page that curates AI-related papers published worldwide.
All content here is summarized using Google Gemini and operated on a non-profit basis.
Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.

CLAIM: Clinically-Guided LGE Augmentation for Realistic and Diverse Myocardial Scar Synthesis and Segmentation

Created by
  • Haebom

Author

Farheen Ramzan, Yusuf Kiberu, Nikesh Jathanna, Shahnaz Jamil-Copley, Richard H. Clayton, Chen Chen

Outline

CLAIM (Clinically-Guided LGE Augmentation for Realistic and Diverse Myocardial Scar Synthesis and Segmentation) is a deep learning-based framework for myocardial scar segmentation from late gadolinium-enhanced (LGE) images of cardiac magnetic resonance imaging (MRI). To address the limited availability of LGE images and the variability of high-quality scar labels, we present a framework for anatomically informed scar generation and segmentation. Its core module, SMILE (Scar Mask generation guided by cLinical knowledgE), conditions a diffusion-based generative model on a clinically adopted AHA 17-segment model to synthesize images with anatomically consistent and spatially diverse scar patterns. In addition, we use a strategy of co-training the scar segmentation network and the generative model to simultaneously improve the realism and segmentation performance of the synthesized scars. Experimental results demonstrate that CLAIM generates anatomically consistent scar patterns and achieves higher Dice similarity to the real scar distribution than the baseline model. It enables controllable and realistic myocardial scar synthesis and has been demonstrated to be useful in subsequent medical imaging work. The code is available at https://github.com/farheenjabeen/CLAIM-Scar-Synthesis .

Takeaways, Limitations

•
Takeaways:
◦
Presenting an effective data augmentation technique to solve the problem of limited LGE image data
◦
Demonstrating the utility of the SMILE module in synthesizing anatomically consistent and diverse myocardial scar patterns
◦
Validation of the effectiveness of a joint learning strategy that simultaneously improves the realism and segmentation accuracy of synthetic scars
◦
Providing high-quality synthetic data for follow-up medical imaging work
•
Limitations:
◦
Limitations of the SMILE module that relies on the AHA 17-segment model (applicability to other anatomical segmentation methods needs to be reviewed)
◦
Additional validation of the realism of the synthesized data is needed (generalization performance evaluation on various clinical datasets is needed)
◦
Additional performance evaluation of downstream medical imaging tasks using synthesized data is needed.
👍
Made with Slashpage