CNN-Based Cell Analysis: From Image to Quantitative Representation

Cédric Allier,Kiran Padmanabhan,Lionel Hervé,William Pierré,Chiara Paviolo,Olivier Cioni,Francesca Andriani,Sophie Morales,Ondrej Mandula

doi:10.3389/fphy.2021.776805

Abstract

We present a novel deep learning-based quantification pipeline for the analysis of cell culture images acquired by lens-free microscopy. The image reconstruction part of the pipeline features a convolutional neural network performing phase unwrapping and accelerating the inverse problem optimization. It allows phase retrieval at the 4K level (3,840 × 2,748 pixels) in 3 s. The analysis part of the pipeline features a suite of convolutional neural networks estimating different cell metrics from the reconstructed image, that is, cell surface area, cell dry mass, cell length, and cell thickness. The networks have been trained to predict quantitative representation of the cell measurements that can be next translated into measurement lists with a local maxima algorithm. In this article, we discuss the performance and limitations of this novel deep learning-based quantification pipeline in comparison with a standard image processing solution. The main advantage brought by this method is the fast processing time, that is, the analysis rate of ∼25.000 cells measurements per second. Although our proof of principle has been established with lens-free microscopy, the approach of using quantitative cell representation in a deep learning framework can be similarly applied to other microscopy techniques.

Highlights

Convolutional neural networks (CNNs) have proven to efficiently process microscope images of cell cultures [1, 2]
To measure the discrepancies between the quantitative representation predicted by a CNN and the reference, we calculated the structural similarity index (SSIM)
To assess the performance of the CNNbased quantification pipeline, the estimated cell measurement values were compared to the values obtained with the standard image processing pipeline

Summary

Introduction

Convolutional neural networks (CNNs) have proven to efficiently process microscope images of cell cultures [1, 2]. A set of CNNs can replace conventional algorithms into image processing pipelines of cell culture analysis. We are studying whether the full image processing pipeline could be efficiently replaced with a single CNN. We demonstrate that single CNNs applied to optical path difference image (OPD) of cell culture obtained with lens-free microscopy [10, 11] deliver results in agreement with standard image processing. We found that the CNNs can generalize well over different conditions of acquisition They are robust over noise, can handle the presence of non-uniform background, and perform well up to a cell concentration of 365 cells/mm. The CNNs applied to these fast reconstructions deliver results in agreement with the reference values

Methods

Results

Discussion

Conclusion