# DeBCR: a sparsity-efficient framework for image enhancement through a deep-learning-based solution to inverse problems

## TL;DR

DeBCR is a deep learning framework that improves microscopy image quality with fewer resources, making advanced imaging more accessible.

## Contribution

A sparsity-efficient neural network and framework for image enhancement with reduced computational demands.

## Key findings

- DeBCR outperforms ten state-of-the-art models in denoising and deconvolution tasks.
- The framework achieves robust performance across various microscopy modalities with fewer parameters.
- DeBCR includes a Python library and a Napari plugin for easy integration into bioimage analysis workflows.

## Abstract

Computational image enhancement for microscopy facilitates cutting-edge biological discovery. While promising, the commonly used deep learning methods are computationally expensive owing to the use of general-purpose architectures, which are inefficient for microscopy data. Here, we propose a sparsity-efficient neural network for image enhancement as a deep representation learning solution to inverse problems in imaging. To maximize accessibility, we developed a framework named DeBCR, consisting of a modular Python library and a user-friendly point-and-click DeBCR plugin for Napari, a popular bioimage analysis tool. We provide a detailed protocol for using the DeBCR as a library and a plugin, including data preparation, training, and inference. We compare the image restoration performance of DeBCR to ten current state-of-the-art models over four publicly available datasets spanning crucial modalities in advanced light microscopy. DeBCR demonstrates more robust performance in denoising and deconvolution tasks across all assessed microscopy modalities while requiring notably fewer parameters than existing models.

Rui Li and colleagues propose a deep learning solution to inverse problems in imaging. Their sparsity-efficient network and software improve image restoration across advanced light microscopy modalities with fewer parameters than existing models.