Partially interpretable image deconvolution framework based on the Richardson-Lucy model.

Area of Science:

• Microscopy
• Image Processing
• Computational Imaging

Background:

• Aberrations in fluorescence microscopy reduce image quality.
• Image deconvolution is crucial for correcting these aberrations.
• Existing deep learning methods lack interpretability and generalizability.

Purpose of the Study:

• To propose a novel Richardson-Lucy (RL) model-driven deconvolution framework.
• To improve both the performance and speed of fluorescence image reconstruction.
• To develop partially interpretable deep learning models for deconvolution.

Main Methods:

• Integration of RL into a deep feature space for enhanced generalizability.
• Development of two novel neural network architectures within the RL framework.
• Implementation of an accelerated backprojector for faster reconstruction.

Main Results:

• The proposed framework significantly improves image quality compared to traditional and CNN-based methods.
• Achieved a five-fold increase in reconstruction speed over classic RL.
• Demonstrated superior generalizability by introducing RL into deep feature space.

Conclusions:

• The novel RL-driven deconvolution framework offers enhanced performance and speed for fluorescence microscopy.
• Partially interpretable neural networks provide a balance between accuracy and understanding.
• This approach effectively addresses image blurring from out-of-focus phenomena and system aberrations.