Figure 1

Overview of the training procedure and implementation of the CNN-based classifier. (A) Cell isolation: A microscopy image of the silicon/glass dispensing chip is shown. The cell suspension enters the chip via the inlet and ~160 pl droplets are ejected from the 40 µm x 40 µm sized nozzle. Images of the nozzle region are captured by the camera of the cell detection microscope in the single-cell printer. A cropped image of each dispensed single-cell with a size of 55 × 55 pixels is stored and can be unambiguously linked to the microwell the cell was deposited into. (B) Cultivation: Colony growth is assessed by imaging the plates. Cells that resulted in viable colonies after 10 days are labeled “viable”, cells that did not grow are labeled “dead”. (C) Training: Labeled cell images are used to train a convolutional neural network (CNN) with one output node for binary classification of input images 55 × 55 pixels in size. The flowchart in (D) illustrates the implementation of the additional CNN-based cell classifier into the cell detection algorithm of the single-cell printer: Only if the feature-based object detection algorithm identified a single cell in the nozzle, a cropped region with the cell is selected for additional prediction on the trained CNN (indicated in blue).