Denoising, Deblurring, and optical Deconvolution for cryo-ET and light microscopy with a physics-informed deep neural network DeBCR

Abstract

Computational image-quality enhancement for microscopy (deblurring, denoising, and optical deconvolution) provides researchers with detailed information on samples. Recent general-purpose deep learning solutions advanced in this task. Yet, without consideration of the underlying physics, they may yield unrealistic and non-existent details and distortions during image restoration, requiring domain expertise to discern true features from artifacts. Furthermore, the large expressive capacity of general-purpose deep learning models requires more resources to train and use in applications. We introduce DeBCR, a physics-informed deep learning model based on wavelet theory to enhance microscopy images. DeBCR is a light model with a fast runtime and without hallucinations. We evaluated the image restoration performance of DeBCR and 12 current state-of-the-art models over 6 datasets spanning crucial modalities in advanced light microscopy and cryo-electron tomography. Leveraging optic models, DeBCR demonstrates superior performance in denoising, optical deconvolution, and deblurring tasks across both LM and cryo-ET modalities.