Kernel learning enables fluorescence microscopic image deconvolution with enhanced performance and speed

The contrast and resolution of fluorescence microscopic images can be effectively improved by classic iterative Richardson-Lucy Deconvolution (RLD) algorithm, but this method is computationally expensive, particularly for three-dimensional data. Variants of RLD with manually designed unmatched backward projector can greatly accelerate deconvolution, however, they require careful parameter optimization to avoid introducing artifacts. Here, we develop Kernel Learning Deconvolution (KLD), which automatically learns forward/backward kernel in RLD from only one paired low-resolution and high-resolution images. The learned kernel reveals a similar pattern with handcrafted Wiener-Butterworth kernel but is more adaptive to data. Besides, it is robust to the signal-to-noise ratio and the number of training samples. KLD shows enhanced deconvolution performance and speed on different cellular structures and imaging modalities, including wide-field microscopy, confocal microscopy, and lattice light-sheet microscopy.