Optimizing deep learning model for measuring 3D dynamics of single RBCs with AI-based digital holographic microscopy

Red blood cells (RBCs) can be important biomarkers for hematologic diseases as their morphology and hemodynamics may change. Current methods to measure the 3D translational and rotational motions of RBCs require complicated optical setups and post-processing, while being limited to straight microchannels. Thus, there is a strong need for a simple and effective method to measure RBCs’ 3D dynamics precisely. This study utilized novel digital in-line holographic microscopy and deep learning techniques to record holograms of RBCs flowing in a viscous fluid. The 3D position of RBCs was found by mathematically reconstructing holograms at various heights to find each cell’s focal plane and determining its in-plane centroid. Holograms of moving RBCs and their respective out-of-plane angle labels were used to train deep learning models. Supervised, residual, and self-supervised models were trained to find the best network for predicting the out-of-plane angle. The in-plane angle of each RBC was measured from the inclined angle of its major axis. The full 3D translational and rotational information of RBCs can be analyzed from sequential holographic images. These methods can be applied to various microfluidic channels to comparatively examine the 3D dynamics of RBCs with a high throughput.