Enhancing 3D Cardiac CT Segmentation with Latent Diffusion Model and Self-Supervised Learning

Quanqi Hu, Ashok Vardhan Addala, Masaki Ikuta, Ravi Soni, Gopal Avinash
Proceedings of the 4th Machine Learning for Health Symposium, PMLR 259:490-501, 2025.

Abstract

CT cardiac imaging remains one of the most challenging visualization techniques among numerous CT organ imaging procedures. This is because of the dynamic nature of human hearts, constantly moving and pumping blood. Due to cardiac motions, CT scanners need to be capable of taking fast scans to capture a “snapshot” of a human heart. Other cardiac imaging challenges include contrast timing variations, radiation dose to patient bodies, limited temporal resolution, contrast agent allergies, and more. In this paper, we present a new latent diffusion model for 3D CT cardiac imaging where the model produces both image volumes and segmentation labels. The latent diffusion model is trained with distinct data augmentation techniques to enhance the variety of the generative data. This helps capture the dynamic nature of the cardiac images. The generative data are used in our Self-Supervised Learning (SSL) to pre-train our Deep Learning (DL) model. Furthermore, because our latent diffusion model produces both images and segmentation labels, our fine-tuning process takes advantage of the diffusion-generated images and labels in addition to the GT data. We run extensive experiments to show that the latent diffusion model and the SSL do help improve 3D CT cardiac image segmentation performance.

Cite this Paper

BibTeX
@InProceedings{pmlr-v259-hu25a, title = {Enhancing 3D Cardiac CT Segmentation with Latent Diffusion Model and Self-Supervised Learning}, author = {Hu, Quanqi and Addala, Ashok Vardhan and Ikuta, Masaki and Soni, Ravi and Avinash, Gopal}, booktitle = {Proceedings of the 4th Machine Learning for Health Symposium}, pages = {490--501}, year = {2025}, editor = {Hegselmann, Stefan and Zhou, Helen and Healey, Elizabeth and Chang, Trenton and Ellington, Caleb and Mhasawade, Vishwali and Tonekaboni, Sana and Argaw, Peniel and Zhang, Haoran}, volume = {259}, series = {Proceedings of Machine Learning Research}, month = {15--16 Dec}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v259/main/assets/hu25a/hu25a.pdf}, url = {https://proceedings.mlr.press/v259/hu25a.html}, abstract = {CT cardiac imaging remains one of the most challenging visualization techniques among numerous CT organ imaging procedures. This is because of the dynamic nature of human hearts, constantly moving and pumping blood. Due to cardiac motions, CT scanners need to be capable of taking fast scans to capture a “snapshot” of a human heart. Other cardiac imaging challenges include contrast timing variations, radiation dose to patient bodies, limited temporal resolution, contrast agent allergies, and more. In this paper, we present a new latent diffusion model for 3D CT cardiac imaging where the model produces both image volumes and segmentation labels. The latent diffusion model is trained with distinct data augmentation techniques to enhance the variety of the generative data. This helps capture the dynamic nature of the cardiac images. The generative data are used in our Self-Supervised Learning (SSL) to pre-train our Deep Learning (DL) model. Furthermore, because our latent diffusion model produces both images and segmentation labels, our fine-tuning process takes advantage of the diffusion-generated images and labels in addition to the GT data. We run extensive experiments to show that the latent diffusion model and the SSL do help improve 3D CT cardiac image segmentation performance.} }
Endnote
%0 Conference Paper %T Enhancing 3D Cardiac CT Segmentation with Latent Diffusion Model and Self-Supervised Learning %A Quanqi Hu %A Ashok Vardhan Addala %A Masaki Ikuta %A Ravi Soni %A Gopal Avinash %B Proceedings of the 4th Machine Learning for Health Symposium %C Proceedings of Machine Learning Research %D 2025 %E Stefan Hegselmann %E Helen Zhou %E Elizabeth Healey %E Trenton Chang %E Caleb Ellington %E Vishwali Mhasawade %E Sana Tonekaboni %E Peniel Argaw %E Haoran Zhang %F pmlr-v259-hu25a %I PMLR %P 490--501 %U https://proceedings.mlr.press/v259/hu25a.html %V 259 %X CT cardiac imaging remains one of the most challenging visualization techniques among numerous CT organ imaging procedures. This is because of the dynamic nature of human hearts, constantly moving and pumping blood. Due to cardiac motions, CT scanners need to be capable of taking fast scans to capture a “snapshot” of a human heart. Other cardiac imaging challenges include contrast timing variations, radiation dose to patient bodies, limited temporal resolution, contrast agent allergies, and more. In this paper, we present a new latent diffusion model for 3D CT cardiac imaging where the model produces both image volumes and segmentation labels. The latent diffusion model is trained with distinct data augmentation techniques to enhance the variety of the generative data. This helps capture the dynamic nature of the cardiac images. The generative data are used in our Self-Supervised Learning (SSL) to pre-train our Deep Learning (DL) model. Furthermore, because our latent diffusion model produces both images and segmentation labels, our fine-tuning process takes advantage of the diffusion-generated images and labels in addition to the GT data. We run extensive experiments to show that the latent diffusion model and the SSL do help improve 3D CT cardiac image segmentation performance.
APA
Hu, Q., Addala, A.V., Ikuta, M., Soni, R. & Avinash, G.. (2025). Enhancing 3D Cardiac CT Segmentation with Latent Diffusion Model and Self-Supervised Learning. Proceedings of the 4th Machine Learning for Health Symposium, in Proceedings of Machine Learning Research 259:490-501 Available from https://proceedings.mlr.press/v259/hu25a.html.

Related Material