T-FDR: Trust-guided frequency-aware diffusion refinement for reliable cardiac MRI segmentation under limited annotations.
Authors
Affiliations (7)
Affiliations (7)
- School of Computer Science and Engineering, Southeast University, No. 2 Southeast University Road, Nanjing, 210096, Jiangsu, China. Electronic address: [email protected].
- College of Software, Nankai University, No. 38 Tongyan Road, Tianjin, 300450, China. Electronic address: [email protected].
- School of Computer Science and Engineering, Southeast University, No. 2 Southeast University Road, Nanjing, 210096, Jiangsu, China. Electronic address: [email protected].
- School of Computer Science and Engineering, Southeast University, No. 2 Southeast University Road, Nanjing, 210096, Jiangsu, China. Electronic address: [email protected].
- School of Computer Science and Engineering, Southeast University, No. 2 Southeast University Road, Nanjing, 210096, Jiangsu, China. Electronic address: [email protected].
- School of Computer Science and Engineering, Southeast University, No. 2 Southeast University Road, Nanjing, 210096, Jiangsu, China. Electronic address: [email protected].
- School of Computer Science and Engineering, Southeast University, No. 2 Southeast University Road, Nanjing, 210096, Jiangsu, China. Electronic address: [email protected].
Abstract
Semi-supervised image segmentation is often limited by noisy pseudo-label propagation and insufficient modeling of boundary-sensitive features and latent semantic distributions, especially in cardiac magnetic resonance imaging (MRI) with low contrast and ambiguous structures. To address these challenges, we propose a trust-guided frequency-aware diffusion refinement (T-FDR) framework that formulates segmentation as a progressive refinement process. Specifically, the framework consists of three sequential stages: suppressing unreliable supervision, recovering boundary-sensitive features, and aligning latent semantic distributions. A Trust-guided Feature Structuring (TFS) module reduces noisy supervision by estimating pseudo-label reliability, followed by a Frequency-aware Residual Refinement (FRRF) module that enhances high-frequency boundary details in uncertain regions. Finally, a Bidirectional Diffusion Correction Module (BDCM) progressively refines latent semantic representations to improve consistency between labeled and unlabeled data. Experiments on three public cardiac MRI benchmarks demonstrate that T-FDR consistently outperforms state-of-the-art methods in both overlap and boundary metrics, indicating its effectiveness for robust semi-supervised segmentation under limited annotations.