Preliminary study for counting fossil diatoms using a deep learning system: An approach to automated estimation of a paleoenvironmental index

Abstract

Two types (intercalary and terminal) of valves of Eucampia antarctica, a species of diatom, have shown potential as paleoenvironmental tools in the Southern Ocean. Taxonomists have counted the valves manually; however, they have required considerable time to assess the relationship between the ratio of the valves and environmental factors. Here, we present an end-to-end automatic approach for counting E. antarctica using the microfossil classification and rapid accumulation device (miCRAD) system, which enables model classification while acquiring microscopic images. We constructed a deep learning-based model for identifying the intercalary and terminal valves of E༎antarctica in a diatom assemblage. Additionally, we tested whether the constructed model functions as a manual count using an experimental image dataset containing all particle images acquired during the whole-scanning of permanent slides. Following cross-validation to verify the model performance potential, the model accuracy reached 0.92 with the use of the training images. The proportion of intercalary valves to all E. antarctica valves (i.e., a total of terminal and intercalary valves) calculated from the model counts yielded 0.55 on average, showing a + 0.05% difference against the actual value of 0.50. However, using the experimental dataset, the model classifications performed worse than the ones estimated based on the cross-validation. The lower performance was attributed to the imbalanced class dataset from the whole-scanning of permanent slides, which includes many other particles. This experiment demonstrated that the classification model constructed with miCRAD system has comparable performance in predicting E.antarctica valves to manual counting; however, screening images before the classification step will be necessary to completely automate the classification.