Bimodal Fusion of Single-Cell Impedance and Imaging Data Using Five Deep Neural Networks

Single-cell electrical and structural properties measured by impedance and images play critical roles in clinical examinations. Although leukocyte differentials based on single-cell impedance or images have been widely studied, bimodal fusion of single-cell impedance and imaging data still lacks a systematic study. Single-cell impedance and images acquired by a home-developed impedance and imaging flow cytometry of microfluidics were merged and processed by a convolutional neural network for data-level direct modeling. Meanwhile, recessive features extracted from impedance and images were fused and classified by a support vector machine and a feedforward neural network for feature-level marginal/joint fusions. Preliminary outputs of classifying single-cell impedance and images were directly averaged or processed by a feedforward neural network for decision-level averaging and meta-learning. Classification accuracies of the four-part leukocyte differential were quantified as 98.9% for data-level direct modeling, 99.3% for feature-level marginal fusion, 99.1% for feature-level joint fusion, 99.6% for decision-level averaging and 99.5% for decision-level meta-learning, which were significantly higher than 90.4% of impedance and 90.2% of images of single cells only. The classification accuracy of four-part leukocyte differential was significantly improved based on single-cell impedance and images coupled with bimodal fusion strategies.