Identification of environmental microorganism using optimally fine-tuned convolutional neural network

Area of Science:

• Environmental microbiology
• Machine learning applications
• Digital microscopy

Background:

• Dense convolutional network with 201 convolutional layers (DenseNet-201) performance can be enhanced through hyper-parameter optimization and data augmentation.
• DenseNet-201 has been rarely applied to environmental microorganism (EM) image identification.
• Accurate classification of EM images is crucial for environmental monitoring and research.

Purpose of the Study:

• To propose an optimally fine-tuned DenseNet-201 (OFTD) model incorporating data augmentation for improved EM image classification.
• To evaluate the performance of the OFTD model on the Environmental Microorganism Dataset (EMDS).
• To visually interpret the feature importance of the OFTD model using gradient-weighted class activation mapping (Grad-CAM).

Main Methods:

• Utilized 70% of the EMDS dataset for training the OFTD model's convolutional layers.
• Employed the remaining EMDS images as a testing dataset to evaluate OFTD performance.
• Applied gradient-weighted class activation mapping (Grad-CAM) for visual feature illustration.

Main Results:

• The OFTD model with data augmentation achieved a classification accuracy of 98.4% on the EMDS dataset.
• Optimally fine-tuning the classification layer proved more effective than data augmentation alone for performance improvement.
• Grad-CAM successfully highlighted key features, such as the foot of Rotifera and stalk of Vorticella.

Conclusions:

• The proposed OFTD model with data augmentation offers a stable and accurate solution for EM detection in digital microscopy.
• Fine-tuning the classification layer is a critical factor in enhancing DenseNet-201 performance for EM image analysis.
• The study demonstrates the potential of deep learning models for automated environmental microorganism identification.