Pipeline for FlowCam data processing with modular open-source software and optional machine learning classification
Abstract
Imaging instruments are becoming widely used in plankton research as they offer several advantages over traditional microscopy: 1) processing orders of magnitude more samples and organisms per time, 2) collection of more quantitative trait data from all organisms, 3) reducing human bias including the possibility to reanalyse image data, 4) rapid imaging of samples avoiding bias by deterioration of preserved samples before analysis, and further 5) some imagers allow analysis of live organisms enabling detection and quantification of delicate organisms that cannot be properly fixed. However, processing the huge number of images produced by common plankton imagers such as the FlowCam (Yokogawa Fluid Imaging Technologies, Inc., ME, USA) remains challenging. VisualSpreadsheet (VSP) - the commercial software necessary to operate FlowCam instruments - offers images, associated particle properties, and statistical analysis tools. However, it has licensing costs, runs exclusively on Windows, and offers limited support for older software versions and machine learning classification. Third-party alternatives to VSP for image sorting and classification have shortcomings related to data format and applicability across systems. We developed a freely available, multi-platform, modular pipeline for processing FlowCam data from various instruments and VSP versions, while also adding important functionalities. A preprocessing Python script unifies the output of different VSP versions and detects imaging artefacts (air bubbles, beads, duplicate images). The size range of target particles can be determined by a user-defined threshold, and their individual biovolume is calculated based on a distance map algorithm. The preprocessed data is summarised in a CSV file that can be opened in LabelChecker, the open-source, cross-platform program presented here. LabelChecker displays FlowCam images without transforming the FlowCam’s output format and enables annotation and validation of labels. The processing pipeline can be paired with machine learning approaches for automatic image classification. Classification results are stored in the same CSV file that opens with LabelChecker for easy label validation and further data processing. We demonstrate the workflow of this pipeline with two plankton datasets. We first annotate images and then use them to train a custom shallow, multi-input classification model. Focussing on accessibility, this pipeline (preprocessing, LabelChecker, and machine learning) paves the way for fast and reproducible plankton analysis by FlowCam, to enable high-throughput analyses adaptable to a wide range of plankton studies. This freely available plankton imaging pipeline can facilitate a wider use of FlowCam instruments and their data, increasing the overall scientific output. Its modular design also makes it easily adaptable to data from virtually any plankton imaging system.