Focus prediction of medical microscopic images based on Lightweight Densely Connected with Squeeze-and-Excitation Network

• Published in: Frontiers in neuroscience 2023-06, Vol.17, p.1213176-1213176
• Main Authors: Jiang, Hesong, Ma, Li, Wang, Xueyuan, Zhang, Juan, Liu, Yueyue, Wang, Dan, Wu, Peihong, Han, Wanfen
• Format: Article
• Language: English
• Subjects: Accuracy; Algorithms; Datasets; Deep learning; DenseNet; focus prediction; Light microscopy; Mathematical models; medical microscopy; Neural networks; Neuroscience; Photobleaching; Phototoxicity; Predictions; squeeze and excitation
• ISSN: 1662-4548; 1662-453X; 1662-453X
• DOI: 10.3389/fnins.2023.1213176

Due to the demand for sample observation, optical microscopy has become an essential tool in the fields of biology and medicine. In addition, it is impossible to maintain the living sample in focus over long-time observation. Rapid focus prediction which involves moving a microscope stage along a vertical axis to find an optimal focus position, is a critical step for high-quality microscopic imaging of specimens. Current focus prediction algorithms, which are time-consuming, cannot support high frame rate imaging of dynamic living samples, and may introduce phototoxicity and photobleaching on the samples. In this paper, we propose Lightweight Densely Connected with Squeeze-and-Excitation Network (LDSE-NET). The results of the focusing algorithm are demonstrated on a public dataset and a self-built dataset. A complete evaluation system was constructed to compare and analyze the effectiveness of LDSE-NET, BotNet, and ResNet50 models in multi-region and multi-multiplier prediction. Experimental results show that LDSE-NET is reduced to 1E-05 of the root mean square error. The accuracy of the predicted focal length of the image is increased by 1 ~ 2 times. Training time is reduced by 33.3%. Moreover, the volume of the model only reaches the KB level, which has the characteristics of being lightweight.