Traditional fluorescence-activated cell sorting (FACS) is a common method to purify or isolate desired cell populations from the bulk cell suspension with high throughput. However, there are many biological applications that aren’t possible on FACS because they require imaging to classify cell types. A major challenge to image-activated cell sorters (IACS) is its robustness to handle various cell shapes and morphology when using hand-crafted image features in real-time. With recent development in machine learning, convolutional neural networks can learn more effective and compact cell image features than traditional feature design techniques. We use these developments to propose and test a cell image classification workflow for real-time:
1. A spatial feature encoder is trained from collected datasets to identify common features of cell images.
2. Such learned features are used to perform Unsupervised Clustering to identify the subpopulations in the sample, from which users can pick the subpopulations to sort.
3. Images from the selected subpopulations are used to refine a pretrained base neural network for Supervised Classification for the current experiment.
4. The refined classifier is then used to make a sorting decision to determine if the input cell should be sorted or not as it travels through the IACS.
We present results using this workflow that performed label-free classification of 5 types of white blood cells with a 98.0% precision and 93.0% recall with a latency of less than 0.4 ms per cell. This accuracy and speed demonstrate that this classification workflow can be used in an Image Activated Cell Sorter.
