Deep learning has demonstrated significant potential in fluorescence microscopy imaging. However, most existing methods primarily enhance channel or spatial features, overlooking the capabilities of kernel space. This limitation restricts the network's ability to recover fine structures and details, especially under noisy or degraded imaging conditions. In this work, the iKUNet-RCAN model, an architecture that integrates kernel and channel attention mechanisms, is proposed. By explicitly capturing kernel space dependencies, the proposed model enhances feature representation with minimal additional computational cost. Experimental results across multiple microscopy modes (confocal, widefield, and two-photon) reveal superior image quality and reconstruction robustness compared with existing methods.
Zhou et al. (Mon,) studied this question.