AI based image analysis of red blood cells in oscillating microchannels

Abstract

The flow dynamics of red blood cells in vivo in blood capillaries and in vitro in microfluidic channels is complex. Cells can obtain different shapes such as discoid, parachute, slipper-like shapes and various intermediate states depending on flow conditions and their viscoelastic properties. We use artificial intelligence based analysis of red blood cells (RBCs) in an oscillating microchannel to distinguish healthy red blood cells from red blood cells treated with formaldehyde to chemically modify their viscoelastic behavior. We used TensorFlow to train and validate a deep learning model and achieved a testing accuracy of over 97%. This method is a first step to a non-invasive, label-free characterization of diseased red blood cells and will be useful for diagnostic purposes in haematology labs. This method provides quantitative data on the number of affected cells based on single cell classification.

Fig. 1. Schematic of the microfluidic setup mounted on an inverted fluorescence microscope: RBCs entering the device and flow into the region of oscillating width and adopt their shape. Videos are taken with a fast camera and recording is triggered by cell passing a photoelectric barrier and rapid analysis in a photomultiplier tube.

Fig. 2. 2 Data processing pipeline. (A) Single frame from a video clip with a red blood cell in the red boundary box touches the “soft trigger” (black line) which then leads to the extraction of the image in the area “crop” (black box). (B) Background of the cropped area from (A) by taking a video frame from the video clip without a red blood cell. (C) Background subtraction of (B) subtracted from (C) and taken the absolute value. (D) TensorFlow layers: “norm”: RGB value normalisation to 0.01, “conv2D”: standard 2D convolutional layer, “max pooling”: standard 2D max pooling, “dense”: standard dense layer. The final layer has two outputs: one for the detection probability of a native red blood cell and one for the chemically modified one.

Fig. 3. TensorFlow training behaviour and results. (A) Training and validation loss for the narrow part of the microchannel and for the wide part (B) of the microchannel. Classification results for the narrow section of the microchannel for (C) native red blood cells and (D) chemically modified red blood cells. Classification results for the wide section of the microchannel for (E) native red blood cells and (F) chemically modified red blood cells. “N”: native, “C”: chemically modified, “p/det”: detection probability.