A super-resolution strategy for mass spectrometry imaging via transfer learning

Tiepeng Liao, Zihao Ren, Zhaoliang Chai, Man Yuan, Chenjian Miao, Junjie Li, Qi Chen, Zhilin Li, Ziyi Wang, Lin Yi, Siyuan Ge, Wenwei Qian, Longfeng Shen, Zilei Wang, Wei Xiong, Hongying Zhu

Abstract: High-spatial-resolution mass spectrometry imaging (HSR-MSI) provides precise spatial information on thousands of biomolecules without labelling across a tissue section. Deep learning methods, trained on large numbers of images, can be used to further improve resolution. However, the limited amount of HSR-MSI data that are publicly available mean that super-resolution reconstruction of images obtained by MSI using deep learning is still a challenge. Here we develop a deep learning framework based on transfer learning called MSI from optical super-resolution (MOSR) that substantially reduces the requirement for sample size. Needing only ten HSR-MSI images, the method transfers knowledge learned from abundant optical images (~15,000) to MSI tasks. Compared with the deep learning model without transfer learning, the MOSR model obtains better image quality with higher peak signal-to-noise ratios and structural similarity index values. It also achieves higher training efficiency and a stronger generalization performance. The MOSR model predicts HSR-MSI images with very small sample size and could transform applications with super-resolution MSI. Mass spectrometry imaging (MSI) can provide important information, but long imaging times are needed to achieve a high spatial resolution, which is why the amount of publicly available high-resolution MSI data for deep learning applications is limited. Liao and colleagues use transfer learning from optical super-resolution images to reduce the amount of MSI data that is needed.