軽量化されたリチャードソン・ルーシーネットワークによる蛍光顕微鏡デコンボリューションの高速化

Fluorescence microscopy contrast and resolution can be greatly improved through deconvolution. Deep unfolding deconvolution networks have recently received significant attention by combining the interpretability of traditional models with enhanced reconstruction efficiency and quality through data-driven learning. However, the proliferation of network layers complicates computation, reduces the interpretability of the model, and requires more training data with increasing network parameters. In this study, we present the accelerated Richardson-Lucy network (ARLN), which incorporates momentum information and the Richardson-Lucy (RL) model into a lightweight neural network. ARLN emphasizes learning knowledge within the image domain and introduces an additive accelerated vector through minimal trainable convolution layers, which are integrated into the RL iterative scheme to accelerate convergence. Both simulations and real experiments demonstrate that, despite its simplicity, ARLN ensures interpretability and achieves comparable or even superior reconstruction performance, especially in small-sample scenarios. It reduces iterations by 6-10 × compared to the accelerated Richardson-Lucy (ARL) algorithm, achieves a greater reduction than 70% in computational cost and the number of parameters relative to lightweight networks, and reconstructs approximately 4-12 × faster than two representative model-driven networks. Owing to its lightweight design and accelerated deconvolution capability, ARLN has great potential and practical importance in restoring fast and resource-constrained fluorescence microscopy.