Joint-embedding self-supervised representation learning for robust cone-beam CT segmentation.
Authors
Affiliations (6)
Affiliations (6)
- Department of Radiation Oncology, Winship Cancer Institute, Emory University School of Medicine, Atlanta, GA, 30322, USA; Department of Radiation Oncology, Taussig Cancer Center, Cleveland Clinic, Cleveland, OH, 44106, USA.
- Nell Hodgson Woodruff School of Nursing, Emory University, Atlanta, GA, 30322, USA.
- Department of Radiation Oncology, Winship Cancer Institute, Emory University School of Medicine, Atlanta, GA, 30322, USA.
- Department of Radiation Oncology, Winship Cancer Institute, Emory University School of Medicine, Atlanta, GA, 30322, USA; School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA, 30332, USA.
- Department of Radiation Oncology, Winship Cancer Institute, Emory University School of Medicine, Atlanta, GA, 30322, USA. Electronic address: [email protected].
- Department of Radiation Oncology, Winship Cancer Institute, Emory University School of Medicine, Atlanta, GA, 30322, USA; School of Mechanical Engineering, Georgia Institute of Technology, Atlanta, GA, 30332, USA; Department of Radiation and Cellular Oncology, The University of Chicago, Chicago, IL, 60637, USA. Electronic address: [email protected].
Abstract
Cone-beam CT (CBCT) segmentation is challenged by reconstruction artifacts and limited reliable annotations. We investigate joint-embedding predictive architecture (JEPA)-style self-supervised learning for volumetric CBCT segmentation using CBCT-LGR, which extends our preliminary slice-wise conference study through fully three-dimensional local-global encoding, learnable semantic tokens, and a combined objective comprising local-global alignment, semantic alignment, and predictive consistency. We compare CBCT-LGR with supervised baselines (nnU-Net and STU-Net), generic self-supervised methods (BYOL-Med3D and Models Genesis), and foundation models (SAM, MedSAM, and Anatomix) using harmonized patient-level splits, preprocessing, and evaluation metrics while retaining each method's stated supervision setting. For representation-learning methods, frozen encoders are evaluated using supervised decoders trained with the same 10-subject labeled-data budget per dataset. Across three datasets, CBCT-LGR outperforms the generic self-supervised baselines on dataset-average overlap and boundary metrics. On CBCTLiTS, supervised and fine-tuned foundation-model approaches achieve higher absolute DSC under clean reconstruction conditions. Across five controlled reconstruction-quality levels, however, CBCT-LGR exhibits the smallest maximum relative DSC decline among all nine methods, with declines of 1.5% for liver and 14.5% for tumor, and achieves the highest numerical tumor DSC at the two most degraded levels. Results on Dataset 1 quantify agreement with imperfect reference annotations and do not establish true anatomical accuracy or a definitive method ranking. These findings define the capabilities and limitations of JEPA-style pretraining for volumetric CBCT segmentation.