Deep Learning Quality Control for High-Throughput High-Content Screening Microscopy Images

Automation of biological image analysis is essential to boost biomedical research. The study of complex diseases such as neurodegenerative diseases calls for big amounts of data to build models towards precision medicine. Such data acquisition is feasible in the context of high-throughput high-content screening (HTHCS) in which the quality of the results relays on the accuracy of image analysis. Deep learning (DL) yields great performance in image analysis tasks especially with big amounts of data such as the produced in HTHCS contexts. Such DL and HTHCS strength is also their biggest weakness since DL solutions are highly sensitive to bad quality datasets. Hence, accurate Quality Control (QC) for microscopy HTHCS becomes an essential step to obtain reliable pipelines for HTHCS analysis.
Usually, artifacts found on these platforms are the consequence of out-of-focus and undesirable density variations. The importance of accurate outlier detection becomes essential for both the training process of generic ML solutions (i.e. segmentation or classification) and the QC of the input data such solution will predict on. Moreover, during the QC of the input dataset, we aim not only to discard unsuitable images but to report the user on the quality of its dataset giving the user the choice to keep or discard the bad images.
To build the QC solution we employed fluorescent microscopy images of rosella biosensor generated in the HTHCS platform. A total of 15 planes ranging from -6z to +7z steps to the two optimum planes. We evaluated 27 known focus measure operators and concluded that they have low sensitivity in noisy conditions. We propose a CNN solution which predicts the focus error based on the distance to the optimal plane, outperforming the evaluated focus operators. This QC allows for better results in cell segmentation models based on U-Net architecture as well as promising improvements in image classification tasks.