Red blood cell classification in lensless single random phase encoding using convolutional neural networks

Rapid cell identification is achieved in a compact and field-portable system employing single random phase encoding to record opto-biological signatures of living biological cells of interest. The lensless, 3D-printed system uses a diffuser to encode the complex amplitude of the sample, then the encoded signal is recorded by a CMOS image sensor for classification. Removal of lenses in this 3D sensing system removes restrictions on the field of view, numerical aperture, and depth of field normally imposed by objective lenses in comparable microscopy systems to enable robust 3D capture of biological volumes. Opto-biological signatures for two classes of animal red blood cells, situated in a microfluidic device, are captured then input into a convolutional neural network for classification, wherein the AlexNet architecture, pretrained on the ImageNet database is used as the deep learning model. Video data was recorded of the opto-biological signatures for multiple samples, then each frame was treated as an input image to the network. The pre-trained network was fine-tuned and evaluated using a dataset of over 36,000 images. The results show improved performance in comparison to a previously studied Random Forest classification model using extracted statistical features from the opto-biological signatures. The system is further compared to and outperforms a similar shearing-based 3D digital holographic microscopy system for cell classification. In addition to improvements in classification performance, the use of convolutional neural networks in this work is further demonstrated to provide improved performance in the presence of noise. Red blood cell identification as presented here, may serve as a key step toward lensless pseudorandom phase encoding applications in rapid disease screening. To the best of our knowledge this is the first report of lensless cell identification in single random phase encoding using convolutional neural networks.