Conditional generative adversarial network-based image enhancement model for confocal microscopy
2026 3rd Computational Optical Imaging and Artificial Intelligence in Biomedical Sciences, California, United States Of America, 17 - 20 January 2026, vol.13865, (Full Text)
- Publication Type: Conference Paper / Full Text
- Volume: 13865
- Doi Number: 10.1117/12.3079178
- City: California
- Country: United States Of America
- Keywords: Artificial intelligence, confocal microscopy, convallaria, deep learning, generative adversarial networks, image processing, zebrafish
- Dokuz Eylül University Affiliated: Yes
Abstract
Confocal laser scanning microscopy provides high-contrast, high-resolution images by precisely scanning the focal plane with single-point detection and advanced illumination techniques. However, this focal-scanning process is inherently time-consuming, leading to long acquisition times. Additionally, exposure to high-intensity laser light can cause photobleaching. High-speed surface-scanning techniques overcome these limitations to some extent by reducing imaging time. While these techniques significantly accelerate image acquisition, they compromise image quality due to subsampling. In this study, we present a reconstruction framework based on a conditional generative adversarial network that restores high-quality confocal images from subsampled data. The proposed model produces images that are substantially more realistic than those generated by existing state-of-the-art methods. A significant concern in generative artificial intelligence is ensuring the accuracy and fidelity of the generated content. The introduction of unrealistic or misleading structures in reconstructed images could lead to incorrect interpretations and conclusions in downstream biological research. To address this issue, our method integrates discrete wavelet transforms within a U-Net-based convolutional neural network architecture. This combination helps preserve authentic structural information while minimizing the likelihood of hallucinated or nonphysical features. We evaluated our reconstruction framework using two distinct biological samples: Convallaria and zebrafish. Experimental results demonstrate that the proposed framework significantly reduces acquisition time while maintaining image quality comparable to that of the original high-quality confocal images.