TS-DENet: a transferable self-supervised learning method for multi-modal fluorescence image denoising

Recent fluorescence diagnostic tools have demonstrated effectiveness in detecting early-stage neoplasmatic tissue and monitoring therapy, allowing rapid non-invasive live imaging diagnosis. However, varying light conditions in in vivo environments and modalities of observation systems introduce multi-level noises to acquired images, causing degraded image quality. Deep learning (DL) has shown great potential in improving image quality, but its performance may be limited when dealing with insufficient labeled training data and the challenges of acquiring high-quality multi-modality fluorescence images in specific biomedical tasks. To address this problem, we propose a two-stage deep denoising and edge enhancement framework (TS-DENet), including large-dataset-based pre-training and domain-specific fine-tuning. The pre-training stage learns contextual features and complex data distribution via a masked reconstruction task. The fine-tuning stage further focuses on denoising and applies edge enhancement to eliminate the image blur induced by denoising. Through extensive experiments, TS-DENet demonstrates state-of-the-art performance in diversified data regimes. Compared with other DL-based methods, TS-DENet shows better generalizability and transferability. For an in vivo experiment, we apply TS-DENet to a multimode fiber (MMF) endoscopic system to observe gastric tissues in a rat. The results suggest that TS-DENet provides a potential solution for fluorescence image quality improvement in clinical practice.