TransFusion-net for multifocus microscopic biomedical image fusion.

Area of Science:

• Biomedical Imaging
• Computer Vision
• Deep Learning

Background:

• Microscopic imaging often suffers from limited depth of focus (DOF), hindering full image clarity.
• Multifocus biomedical image fusion (MFBIF) is a technique to overcome DOF limitations.
• Existing methods may not fully leverage information richness and visual authenticity.

Purpose of the Study:

• To propose a novel transformer network, TransFusion-Net, for multifocus biomedical image fusion.
• To improve image clarity and detail in high-magnification microscopy.
• To achieve accurate and visually authentic fused images.

Main Methods:

• TransFusion-Net employs an interlayer cross-attention module to capture long-range dependencies among non-focus source images.
• A spatial attention upsampling network (SAU-Net) module is utilized to obtain global semantic information.
• The network processes multiple input images in an end-to-end manner, exploiting correlations between source images.

Main Results:

• TransFusion-Net outperformed eight state-of-the-art algorithms in quantitative evaluations using metrics like Q^AB/F, Q_MI, Q^AVG, Q_CB, and PSNR.
• Qualitative experiments demonstrated near-zero residuals, validating the method's fusion adequacy.
• The proposed method consistently produced high-quality fusion results for multifocus biomedical microscopy images.

Conclusions:

• A deep convolutional neural network incorporating joint cross-attention and spatial attention mechanisms can achieve accurate and effective multifocus biomedical microscopic image fusion.
• TransFusion-Net offers a reliable solution for enhancing clarity in challenging microscopic imaging scenarios.
• The approach ensures both information richness and visual authenticity in the fused images.