Automated lensless blood sample identification through scattering media using deep learning architectures

Lensless devices paired with deep learning models have recently shown great promise as a novel approach to biological screening. As a first step toward performing automated lensless cell identification non-invasively, we present a field-portable, compact lensless system that can detect and classify smeared whole blood samples through layers of scattering media. In this system, light from a partially coherent laser diode propagates through the sample, which is positioned between two layers of scattering media, and the resultant opto-biological signature is captured by an image sensor. The signature is transformed via local binary pattern (LBP) transformation, and the resultant LBP images are processed by a convolutional neural network (CNN) to identify the type of red blood cells in the sample. We validated our system in an experimental setup where whole blood samples are placed between two diffusive layers of increasing thickness, and the robustness of the system against variations in the layer thickness is investigated. Several CNN models were considered (i.e., AlexNet, VGG-16, and SqueezeNet), individually optimized, and compared against a traditional learning model that consists of principal component decomposition and support vector machine (PCA + SVM). We found that a two-stage SqueezeNet architecture and VGG-16 provide the highest classification accuracy and Matthew's correlation coefficient (MCC) score when applied to images acquired by our lensless system, with SqueezeNet outperforming the other classifiers when the thickness of the scattering layer is the same in training and test data (accuracy: 97.2%; MCC: 0.96), and VGG-16 resulting the most robust option as the thickness of the scattering layers in test data increases up to three times the value used during training. Altogether, this work provides proof-of-concept for non-invasive blood sample identification through scattering media with lensless devices using deep learning. Our system has the potential to be a viable diagnosis device because of its low cost, field portability, and high identification accuracy.