DFMF: Harnessing spectral-spatial synergy for MR image segmentation through Dual-Task Feature Mining Framework.

Automated segmentation of Magnetic Resonance (MR) images plays a critical role in medical applications, including tumor delineation, organ volume measurement, and lesion tracking. While traditional supervised learning methods depend heavily on costly annotated data, MR images inherently contain rich anatomical information, such as the shape, size, and spatial relationships of organs and tissues. Effectively leveraging this information to enhance segmentation performance remains a significant challenge in current research. To address this, we propose a novel Dual-task Feature Mining Framework (DFMF), which integrates self-supervised and semi-supervised learning paradigms. DFMF simultaneously optimizes two complementary tasks: image inpainting and segmentation, enabling the extraction of richer and more discriminative feature representations. This dual-task mechanism enhances the model's ability to capture complex anatomical structures, leading to superior segmentation performance. To maximize the utility of unannotated data, we introduce a Self-consistency Loss, which enforces consistency between inpainted and original images without requiring explicit data augmentation. Additionally, we design a Hybrid Receptive Field Network (HRFNet) as the backbone of DFMF, which effectively captures global frequency-domain information while preserving fine spatial details. Extensive experiments on four MR image datasets demonstrate that DFMF outperforms state-of-the-art segmentation methods, and ablation studies validate the contribution of each component from multiple perspectives.