Deep learning-powered virtual fluorescent labeling of vascular structures in engineered tissues

Introduction::
Capillary morphogenesis and angiogenesis are crucial to the fields of therapeutic implantable devices, tumor growth, and metastasis. Tissue engineering models are often used as investigational tools in these fields. Here we describe a deep learning model for imaging such tissues without the use of fluorescent labels. We examined a 3-D fibrin hydrogel co-culture system comprising vascular endothelial cells (ECs) and stromal cells (normal human dermal fibroblasts), which are required to induce spontaneous formation of capillary networks. One considerable challenge is the need to fluorescently label cells in order to identify cell type and developing structures over days or weeks. Common labeling techniques include (1) chemical fixation with fluorescence labeling that terminates each experiment timeline, (2) genetic modification to express fluorescent proteins which may alter cell phenotype, and (3) use of cell-tracking dyes that are diluted by half with each cell division. Recently, deep learning models have been used to predict labeled images from one or more unlabeled input images. Here we investigate deep learning-powered tool for predicting fluorescently labeled optical sections from a 3-D culture. Our model utilizes laser scanning reflection confocal microscopy (RCM) images at three laser wavelengths as the input and fluorescently labeled nuclei, actin cytoskeleton and vessel lumen as the ground truth (Fig. 1A). We hypothesis that our trained model will be able to predict fluorescent labels from reflection confocal images to assess cell phenotype and targeted structures during the growth of capillaries, in vitro.




Materials and Methods::
Cell Culture




A 3-D model of capillary morphogenesis was established by co-culturing human umbilical vein endothelial cells (HUVECs) and normal dermal fibroblasts at a 1:1 ratio within 2.5 mg/mL fibrin hydrogels.




Data Acquisition




Our workflow was as follows: image acquisition - > image pre-processing - > deep learning network training - > validation. Our supervised training technique utilizes matched input RCM and ground truth fluorescence confocal microscopy (FCM) images as illustrated in Fig. 1B. For each focal plane, two sets of confocal images were acquired using an Olympus FV3000 laser scanning confocal microscope and an Olympus 40x silicon oil immersion objective lens (NA = 1.25). The first set comprises three RCM images at laser wavelengths 405, 488, and 561 nm. The second set comprises three FCM images of DAPI, phalloidin-Alexa Fluor 488 and lectin UEA I-rhodamine to label nuclei, actin cytoskeletons, and vessel lumens, respectively. Image scan area was 318.20 x 318.20 µm² with a resolution of 4096 x 4096 pixels². Raw images were downscaled to 1024 x 1024 pixels² to reduce computational overhead, normalized to values between 0 and 1, and patched into smaller images of size 512 x 512 pixels².




Network Architecture




One of the most widely used frameworks to perform image-to-image mapping is the U-Net model. A variant of a U-Net model was implemented and optimized using the mean square error (MSE) loss function. Training was performed on 496 patches of size 512 x 512 pixels² and validation was performed on 24 patches of the same size.




Results, Conclusions, and Discussions::
Qualitative examination shows good agreement between predicted and ground truth images (Fig. 1C). Quantitative examination shows an average MSE of 0.3%, 0.9%, and 0.7% for DAPI, phalloidin-Alexa 488 and lectin, respectively. Despite our small sample size, the model performed well at predicting labeled targets with a less than 1% error rate. Current work in our group includes the use of a larger dataset with greater diversity in terms of laser power, image signal over noise, cell density, and image resolution with the aim of increasing model robustness between microscopes and users.




Our deep learning-driven virtual staining technique allows for multi-day or longer longitudinal confocal imaging of capillary morphogenesis within an engineered tissue. But importantly, the method should be generalizable to other cell-cell and cell target systems, which likely will require retraining of this network. Training and implementation do not require a large computer cluster and thus can be replicated in most laboratory setting. Further, nearly all commercial laser scanning microscopes have a reflection confocal microscopy setting thus allowing use of our model. Our Code will be posted on GitHub and made available to the public at no cost. In summary, we provide a new tool for observing tissue formation within transparent engineered tissues without the need for exogenous fluorescent labeling.




Acknowledgements (Optional): :

References (Optional): :