Deep learning-based diatom taxonomy on virtual slides.

Area of Science:

• Algology
• Machine Learning
• Bioinformatics

Background:

• Deep convolutional neural networks (CNNs) are state-of-the-art for image classification.
• Taxonomic identification requires customized deep learning solutions due to diverse organismal groups and imaging techniques.
• Diatoms, a morphologically diverse microalgal group, present unique challenges for automated identification.

Purpose of the Study:

• To demonstrate the application of CNNs for the taxonomic identification of diatoms.
• To investigate factors influencing CNN performance, including architecture, background masking, dataset size, and concept drift.
• To assess the feasibility of domain adaptation for novel taxonomic groups using limited data.

Main Methods:

• Assembled a diatom image database using high-resolution microscopy and web-based annotation from Southern Ocean expeditions.
• Applied and evaluated various CNN architectures, including VGG16, with and without background masking.
• Investigated the impact of dataset size and pre-trained models on classification performance.

Main Results:

• The VGG16 architecture exhibited the best performance and generalization ability for diatom image classification.
• Background masking slightly improved classification performance, contrary to some previous studies.
• Training a classifier on pre-trained convolutional layers achieved high performance (F1 scores ~97%) with only 100-300 examples per class.

Conclusions:

• CNNs, particularly older architectures like VGG16, are effective for diatom taxonomic identification.
• Domain adaptation using pre-trained models is a feasible approach for novel taxonomic groups, requiring minimal effort.
• Customized deep learning solutions are crucial for accurate image-based taxonomic identification across diverse biological groups.