Retaining spatial resolution multifocal confocal fluorescence microscopy with deep learning

Confocal microscopy is a standard modality for volumetric imaging of biological samples due to its high spatial resolution and signal-to-noise ratio (SNR). However, the slow point-by-point scanning process limits its image acquisition speed. Multifocal illumination allows for faster acquisition but compromises spatial resolution. Here, we introduce a deep learning approach for multifocal confocal microscopy that achieves faster acquisition while preserving high resolution. The proposed model is based on image-to-image translation, implemented using modified U-Net, ResU-Net, and Attention U-Net architectures. The model is trained and tested on paired experimental datasets, with conventional confocal images as groundtruth and multifocal confocal images as input from various biological samples. The modified Attention U-Net significantly improves image quality and retains structural details, with higher peak SNR (32.83 dB) and structural similarity index measure (0.935) values. Additionally, spatial frequency analysis and Fourier ring correlation confirm that the Attention U-Net outperforms other models in preserving both low-frequency (>0.92 accuracy) and high-frequency information (0.90 vs. 0.83 for U-Net). Performance metrics demonstrate that our models match the quality of traditional confocal imaging, increasing imaging speed and addressing the trade-off between speed and resolution in multifocal confocal microscopy. These findings underscore the potential of combining deep learning with various confocal imaging applications.