Purification of white blood cells (WBC) and training of WBC classification algorithms for complete blood counts on a new diagnostic platform using lens-free imaging

miDiagnostics is developing a silicon-based nanoFluidic Processor (nFP) embedded in a plastic test card for read-out in a simple lens-free imaging (LFI) system to provide rapid low-cost diagnostic tests of laboratory quality for emergency and/or low-tech settings (Fig1). The nFP miniaturizes a complete diagnostic workflow required to count and differentiate blood cells using solely capillary forces. This poster focuses on another unique aspect of the technology: applying machine learning algorithms to differentiate white blood cells (WBC) (Fig2). Different WBC types were purified from venous whole blood and reference concentration and purification data were obtained with a calibrated haematology analyzer (Sysmex XN350). Negative selection kits from Miltenyi Biotec were used which contained magnetic beads functionalized with antibodies against a variety of surface markers specific for certain blood cell types. As such, it was possible to purify neutrophils and T cells with ≥ 95% purity and B cells, eosinophils and monocytes with ≥ 90% purity. Purification of monocytes required prior density centrifugation using Ficoll to achieve high purity (Fig3). Purified WBC from 10 different donors were imaged by LFI in test cards to obtain reference databases of each cell type (Fig4). The LFI imaging equipment consists of a laser diode light source and a complementary metal-oxide-semiconductor (CMOS) sensor without any lenses, objectives or other optical components. Using algorithmic post-processing, the images captured with the LFI setup were reconstructed from holograms into regular images and were collected in a reference database. A convolutional neural network (CNN) was trained to predict the WBC type based on this reference database, which was split in cell type-balanced training, validation and test sets. The CNN classifier was trained using cross-validation on the training and validation sets. Classifier performance was evaluated against the test set to verify generalization power towards unseen data, showing between 74% and 97% correct classifications (Fig4). In conclusion, immunological purification by negative selection was applied to construct areference database for the development of a WBC classification algorithm. The performance of the algorithm is expected to improve as more data become available. The next step is to assess variability in different donors and different target populations and incorporate this variability into the reference database.