基于生成对抗网络的细胞图像超分辨重建-《中国医学物理学杂志》

文章信息/Info

Title:: Super-resolution reconstruction of cell images based on generative adversarial network

作者:: 邢雄飞1; 樊星佳1; 李睿2; 1.甘肃中医药大学医学信息工程学院，甘肃兰州 730000； 2.兰州大学第二医院信息中心，甘肃兰州730000

Author(s):: XING Xiongfei1; FAN Xingjia1; LI Rui2; 1. School of Medical Information Engineering, Gansu University of Chinese Medicine, Lanzhou 730000, China 2. Information Center, Lanzhou University Second Hospital, Lanzhou 730000, China

Keywords:: Keywords: fluorescence microscopic cell image super-resolution reconstruction generative adversarial network

摘要:: 为提升细胞荧光显微图像的分辨率和重建质量，提出一种基于生成对抗网络的超分辨重建方法。该模型融合卷积与多头自注意力机制，实现多尺度特征建模与融合。通过联合内容损失、感知损失、对抗损失和总变分损失，提升训练稳定性与重建性能。在4种不同细胞亚结构数据集上的实验结果表明，该方法在峰值信噪比和结构相似性指数上均优于现有方法。与次优方法相比，峰值信噪比分别提升0.22、2.35、2.11和0.59 dB，结构相似性指数分别提升0.021、0.009、0.009和0.022。主观视觉评价进一步验证了该方法在细节恢复与伪影抑制方面的优势。

Abstract:: Abstract: To improve the resolution and reconstruction quality of fluorescence microscopic cell images, a super-resolution reconstruction method based on a generative adversarial network is proposed. The model integrates convolutional operations with a multi-head self-attention mechanism to achieve multi-scale feature modeling and fusion. A joint loss function combining content loss, perceptual loss, adversarial loss, and total variation loss is adopted to enhance training stability and reconstruction performance. Experimental results on 4 different cellular substructure datasets demonstrate that the proposed method outperforms existing approaches in terms of peak signal-to-noise ratio and structural similarity index. Compared with the second-best method, the proposed method achieves peak signal-to-noise ratio improvements of 0.22, 2.35, 2.11, and 0.59 dB, as well as structural similarity enhancements of 0.021, 0.009, 0.009, and 0.022, respectively. Subjective visual evaluations further verify the superiority of the proposed method in detail restoration and artifact suppression.