A Transformer Network Based on Taylor Expansion for Generating Phase Contrast Images of Adherent Cells

Yongquan Guo, Jintu Zheng, Yi Ding, Ying Hu, Yi Cao, Zenan Wang

Abstract: The use of automation technology is essential in stem cell research and large-scale preparation to address core technological challenges. The efficiency and accuracy of the observation phase in the large-scale preparation process of stem cells are decisive factors that affect production quality and efficiency. However, existing observation technologies such as phase contrast imaging and traditional computer vision-based methods are inefficient and require manual correction. Therefore, developing a technology that can quickly and efficiently observe the stem cell growth states during the automation process is necessary. A new study introduces a multi-scale Transformer network based on Taylor expansion (TTNet) that aims to monitor stem cell growth states accurately and in real-time during the automation process. This network utilizes a Taylor series expansion to provide an approximation of the conventional softmax attention mechanism and corrects errors in the Taylor expansion through an attention refinement module. Additionally, the network incorporates multi-scale patch embedding technology, achieving effective feature embedding through overlapping deformable convolutions of different scales. By optimizing and compressing the model, this study significantly improves the generation speed of the model without sacrificing the quality of generation. Experimental results demonstrate that TTNet has significant advantages in comprehensive performance over existing generative models in the task of generating bright-field images of stem cells. It achieves the shortest generation time while ensuring the quality of generation.