Fast Multi-Dimensional Imaging Using the Unsupervised 3D Noise2Void Denoising Network.

Area of Science:

• Biological imaging
• Computational imaging
• Machine learning for science

Background:

• Label-free multidimensional imaging techniques like Raman and phase imaging are crucial in biology.
• Raman signals are weak, and phase imaging speed is limited by noise, impacting data quality.
• Existing deep learning denoising methods often require high-SNR data and neglect 3D correlations.

Purpose of the Study:

• To propose and validate an unsupervised deep learning denoising method for label-free multidimensional imaging.
• To address limitations of existing methods by incorporating 3D information and reducing reliance on high-SNR training data.

Main Methods:

• Developed a 3D Noise2Void (3D N2V) network for unsupervised denoising.
• Applied the 3D N2V method to Raman hyperspectral imaging and 3D phase imaging data.
• Compared 3D N2V performance against BM3D and 3D RCAN methods.

Main Results:

• The 3D N2V network effectively removed noise from Raman and phase imaging data in an unsupervised manner.
• The method preserved spectral, axial, and temporal correlations, unlike slice-by-slice approaches.
• 3D N2V demonstrated superior denoising performance, improved limit of detection, and preserved biological features compared to BM3D and 3D RCAN.

Conclusions:

• The proposed 3D N2V method offers an effective unsupervised solution for denoising label-free multidimensional imaging data.
• This technique enhances data quality and biological feature preservation in Raman and phase imaging.
• 3D N2V outperforms existing methods and holds potential for advancing biological imaging analysis.