Deep learning-assisted label-free high-resolution imaging of chromatin in live cell nucle

Ching-Ya Cheng^1*, Yi-Teng Hsiao¹, Huan-Hsin Tseng², Yu Tsao², Chia-Lung Hsieh¹

¹aInstitute of Atomic and Molecular Sciences, Academia Sinica, Taipei, Taiwan
²bResearch Center for Information Technology Innovation, Academia Sinica, Taipei, Taiwan

* Presenter:Ching-Ya Cheng, email:samu172004@gmail.com

Label-free scattering-based interferometric optical microscopy plays a significant role in bioimaging. The stable and indefinite scattering signal enables high-speed, noninvasive observation over a long period of time. However, because linear scattering is ubiquitous, label-free scattering-based imaging lacks specificity against different cell organelles or biomolecules. Recent studies sought to achieve organelle-specific label-free imaging by using the structural information encoded in the phase-sensitive interferograms. Unfortunately, this approach faces limitations in distinguishing organelles having similar sizes and refractive indexes. Here we show a strategy that exploits the dynamic scattering signals of the native cell samples to infer molecularly specific cell images. Even though different cell organelles and biomolecules may exhibit similar structures, the dynamics of these structures would be different due to the distinct interactions with the surrounding environments. Therefore, the addition of temporal information helps to enhance the specificity of label-free imaging. We demonstrated our technique by recording interferograms of live cell nuclei at high speed and calculated the temporal feature maps based on the video. The temporal fluctuation of the signal contains information closely connected to the local mass density of chromatin (the DNA-protein complexes). We use deep learning algorithms (Efficient-Unet) to establish the connections between label-free temporal feature maps and the chromatin fluorescence image. High-fidelity, high-resolution chromatin images are successfully inferred based on the label-free image data set. We examine the impacts of image data quality and machine learning models on overall performance. Our work shows that, in cell imaging applications, temporal information is as useful as structural information. The combination of temporal and structural information is expected to support future investigations of complex cell samples in a label-free manner.

Keywords: Deep learning, label-free, linear scattering, interference microscopy, cell nucleus