Skip to content
Open access

CoBiDiff: anatomy-aware bidirectional diffusion for left atrium cross-modal augmentation

Oct 2026 · Measurement science and technology · Vol 37 · 0 citations · 29 references
Physics

Abstract

Accurate delineation of the left atrium (LA) from cardiac magnetic resonance (MR) and computed tomography (CT) images plays an important role in atrial fibrillation assessment and treatment planning. However, the limited availability of expert-annotated cardiac datasets restricts the robustness and generalizability of deep learning–based segmentation models, particularly in cross-modal scenarios. To overcome this limitation, we propose CoBiDiff, a collaborative bidirectional diffusion framework that explicitly incorporates anatomical constraints into unpaired cross-modal cardiac image augmentation. The framework is developed using unpaired CT and MR volumes from the MICCAI 2017 multi-modality whole heart segmentation (MM-WHS) challenge dataset, focusing specifically on the LA region. A shared diffusion backbone is employed to jointly model both imaging modalities, while signed distance functions (SDFs) are incorporated as structural representations to provide anatomical guidance during generation. To reduce the influence of artificial CT field-of-view boundaries, boundary regions are suppressed through noise replacement and gradient smoothing strategies. In addition, a soft spatial extent loss is introduced to maintain anatomically plausible foreground spatial extents during synthesis. An SDF dropout strategy is further introduced during training to strengthen image-conditioned anatomical inference and improve anatomical coherence. Experimental results indicate that the proposed framework generates LA images with improved structural consistency across modalities. The proposed method achieved Dice coefficients of 86.53% and 95.04%, and average symmetric surface distances of 1.511 mm and 0.623 mm for CT-to-MR and MR-to-CT augmentation, respectively, on the MM-WHS 2017 dataset. CoBiDiff provides a practical framework for expanding limited cardiac imaging datasets without requiring paired cross-modal images. By combining diffusion-based synthesis with anatomy-aware constraints, CoBiDiff improves the downstream utility of generated samples for cardiac segmentation. The full implementation is publicly available at https://github.com/PencilSC/CoBiDiff to support reproducible and open research.

Read PDF

We use cookies to run the site and, with your consent, for analytics and to show ads. See our Cookie Policy.