Curriculum Self-Supervised Learning for 3D CT Cardiac Image Segmentation

Mohammad Reza Hosseinzadeh Taher, Masaki Ikuta, Ravi Soni
Proceedings of the 3rd Machine Learning for Health Symposium, PMLR 225:145-156, 2023.

Abstract

Automating the segmentation of various cardiac chamber structures (e.g., pulmonary artery, aorta, etc.) in 3D CT cardiac imaging remains a significant challenge. This challenge primarily arises from the dynamic nature of the human heart and substantial anatomical variations in terms of organ texture, shape, and size across different patients. These factors collectively result in a scarcity of annotated data, posing a significant hurdle for training data-hungry deep models. The self-supervised learning (SSL) paradigm offers a promising solution to overcome this obstacle since it eliminates the reliance on massive annotated data for training deep models. However, existing SSL approaches fall short in capturing effective representations from 3D cardiac volumes due to the oversight of the dynamic nature of human hearts in the design of their pretext tasks. To address this challenge, we propose a novel SSL method based on the curriculum learning paradigm, which progressively increases the task difficulty during the pretraining stages. Our method enables the SSL model to initially acquire fundamental knowledge about the data, which can subsequently serve as valuable contextual clues for solving more complex tasks during later stages of pretraining. Our extensive experiments demonstrate that the SSL pre-trained model, trained using our strategy, acquires generalizable representations capable of effectively segmenting various existing cardiac chamber structures.

Cite this Paper

BibTeX
@InProceedings{pmlr-v225-taher23a, title = {Curriculum Self-Supervised Learning for 3D CT Cardiac Image Segmentation}, author = {Taher, Mohammad Reza Hosseinzadeh and Ikuta, Masaki and Soni, Ravi}, booktitle = {Proceedings of the 3rd Machine Learning for Health Symposium}, pages = {145--156}, year = {2023}, editor = {Hegselmann, Stefan and Parziale, Antonio and Shanmugam, Divya and Tang, Shengpu and Asiedu, Mercy Nyamewaa and Chang, Serina and Hartvigsen, Tom and Singh, Harvineet}, volume = {225}, series = {Proceedings of Machine Learning Research}, month = {10 Dec}, publisher = {PMLR}, pdf = {https://proceedings.mlr.press/v225/taher23a/taher23a.pdf}, url = {https://proceedings.mlr.press/v225/taher23a.html}, abstract = {Automating the segmentation of various cardiac chamber structures (e.g., pulmonary artery, aorta, etc.) in 3D CT cardiac imaging remains a significant challenge. This challenge primarily arises from the dynamic nature of the human heart and substantial anatomical variations in terms of organ texture, shape, and size across different patients. These factors collectively result in a scarcity of annotated data, posing a significant hurdle for training data-hungry deep models. The self-supervised learning (SSL) paradigm offers a promising solution to overcome this obstacle since it eliminates the reliance on massive annotated data for training deep models. However, existing SSL approaches fall short in capturing effective representations from 3D cardiac volumes due to the oversight of the dynamic nature of human hearts in the design of their pretext tasks. To address this challenge, we propose a novel SSL method based on the curriculum learning paradigm, which progressively increases the task difficulty during the pretraining stages. Our method enables the SSL model to initially acquire fundamental knowledge about the data, which can subsequently serve as valuable contextual clues for solving more complex tasks during later stages of pretraining. Our extensive experiments demonstrate that the SSL pre-trained model, trained using our strategy, acquires generalizable representations capable of effectively segmenting various existing cardiac chamber structures.} }
Endnote
%0 Conference Paper %T Curriculum Self-Supervised Learning for 3D CT Cardiac Image Segmentation %A Mohammad Reza Hosseinzadeh Taher %A Masaki Ikuta %A Ravi Soni %B Proceedings of the 3rd Machine Learning for Health Symposium %C Proceedings of Machine Learning Research %D 2023 %E Stefan Hegselmann %E Antonio Parziale %E Divya Shanmugam %E Shengpu Tang %E Mercy Nyamewaa Asiedu %E Serina Chang %E Tom Hartvigsen %E Harvineet Singh %F pmlr-v225-taher23a %I PMLR %P 145--156 %U https://proceedings.mlr.press/v225/taher23a.html %V 225 %X Automating the segmentation of various cardiac chamber structures (e.g., pulmonary artery, aorta, etc.) in 3D CT cardiac imaging remains a significant challenge. This challenge primarily arises from the dynamic nature of the human heart and substantial anatomical variations in terms of organ texture, shape, and size across different patients. These factors collectively result in a scarcity of annotated data, posing a significant hurdle for training data-hungry deep models. The self-supervised learning (SSL) paradigm offers a promising solution to overcome this obstacle since it eliminates the reliance on massive annotated data for training deep models. However, existing SSL approaches fall short in capturing effective representations from 3D cardiac volumes due to the oversight of the dynamic nature of human hearts in the design of their pretext tasks. To address this challenge, we propose a novel SSL method based on the curriculum learning paradigm, which progressively increases the task difficulty during the pretraining stages. Our method enables the SSL model to initially acquire fundamental knowledge about the data, which can subsequently serve as valuable contextual clues for solving more complex tasks during later stages of pretraining. Our extensive experiments demonstrate that the SSL pre-trained model, trained using our strategy, acquires generalizable representations capable of effectively segmenting various existing cardiac chamber structures.
APA
Taher, M.R.H., Ikuta, M. & Soni, R.. (2023). Curriculum Self-Supervised Learning for 3D CT Cardiac Image Segmentation. Proceedings of the 3rd Machine Learning for Health Symposium, in Proceedings of Machine Learning Research 225:145-156 Available from https://proceedings.mlr.press/v225/taher23a.html.

Related Material