Advancing Medical Image Segmentation with Self-Supervised Learning: A 3D Student-Teacher Approach for Cardiac and Neurological Imaging

Moona Mazher, Daniel C. Alexander, Abdul Qayyum, Steven A Niederer

Keywords: Self-Supervised Learning (SSL), Whole Heart Segmentation (WHS), Ischemic Stroke Lesion Segmentation (ISLES), CT Imaging, MRI Imaging, Cardiac Imaging, Neurological Imaging, xLSTM, Multi-Modal Imaging, Traumatic Brain Injury (TBI).

TL;DR: Introducing a self-supervised 3D student-teacher framework with xLSTM for improved cardiac and neurological image segmentation, addressing data scarcity and modality variability to enhance clinical decision-making.

Abstract: We propose 3D-SegSync, a self-supervised learning (SSL) framework designed to improve segmentation accuracy for both cardiac and neurological structures. It integrates a student-teacher model with a 3D Vision-LSTM (xLSTM) backbone to capture spatial dependencies in volumetric data. The SSL phase utilizes large-scale unlabeled datasets for pretraining, followed by fine-tuning on labeled data to improve segmentation across CT and MRI scans. Experimental results demonstrate that 3D-SegSync achieves consistent performance across different anatomical structures. Additionally, its ability to generalize between CT and MRI without requiring modality-specific modifications highlights its adaptability for cardiac and neurological image segmentation. Given its strong performance, 3D-SegSync has the potential to be extended to other medical image segmentation tasks in the future. Code can be found here: https://github.com/Moona-Mazher/3D-SegSync_SSL.

Primary Subject Area: Segmentation

Secondary Subject Area: Unsupervised Learning and Representation Learning

Paper Type: Both

Registration Requirement: Yes

Reproducibility: https://github.com/Moona-Mazher/3D-SegSync SSL

Visa & Travel: Yes

Midl Latex Submission Checklist: Ensure no LaTeX errors during compilation., Created a single midl25_NNN.zip file with midl25_NNN.tex, midl25_NNN.bib, all necessary figures and files., Includes \documentclass{midl}, \jmlryear{2025}, \jmlrworkshop, \jmlrvolume, \editors, and correct \bibliography command., Did not override options of the hyperref package, Did not use the times package., All authors and co-authors are correctly listed with proper spelling and avoid Unicode characters., Author and institution details are de-anonymized where needed. All author names, affiliations, and paper title are correctly spelled and capitalized in the biography section., References must use the .bib file. Did not override the bibliographystyle defined in midl.cls. Did not use \begin{thebibliography} directly to insert references., Tables and figures do not overflow margins; avoid using \scalebox; used \resizebox when needed., Included all necessary figures and removed *unused* files in the zip archive., Removed special formatting, visual annotations, and highlights used during rebuttal., All special characters in the paper and .bib file use LaTeX commands (e.g., \'e for é)., Appendices and supplementary material are included in the same PDF after references., Main paper does not exceed 9 pages; acknowledgements, references, and appendix start on page 10 or later.

Submission Number: 77