Comparison of Deep Learning Architectures for three different Multispectral Imaging Flow Cytometry Datasets

Abstract

Multispectral imaging flow cytometry (MIFC) is capable of capturing thousands of microscopic multispectral cell images per second. Deep Learning Algorithms in combination with MIFC are currently applied in different areas such as classifying blood cell morphologies, phytoplankton cells of water samples or pollen from air samples or pollinators. The goal of this work is to train classifiers for automatic and fast processing of new samples to avoid labor-intensive and error-prone manual gating and analyses and to ensure rigor of the results. In this study we compare state of the art Deep Learning architectures for the use case of multispectral image classification on datasets from three different domains to determine whether there is a suitable architecture for all applications or if a domain-specific architecture is required. Experiments have shown that there are multiple Convolutional Neural Network (CNN) architectures that show comparable results with regard to the evaluation criteria accuracy and computational effort. A single architecture that outperforms other architectures in all three domains could not be found.

Advancing precision single-cell analysis of red blood cells through semi-supervised deep learning using database of patients with post-COVID-19 syndrome

“…Although MobileNetV2, the previous generation, is reported to have the lowest performance on RBC IFC datasets 20 , the difference is in the margin of 1-2%, and the number of parameters is <10x-100x compared to other popular CNN model families (e.g., ResNet, VGG, Inception, DenseNet). The number of parameters is even smaller with varying width multiplier parameters, allowing rapid training iterations.…”