Comparison among Four Deep Learning Image Classification Algorithms in AI-based Diatom Test.

Area of Science:

• Forensic Science
• Computational Biology
• Digital Pathology

Background:

• Automatic diatom recognition is crucial for forensic investigations, particularly in drowning cases.
• Deep learning offers potential for improving the accuracy and efficiency of diatom testing.

Purpose of the Study:

• To evaluate four deep learning algorithms for automatic diatom recognition.
• To identify the most suitable algorithm for forensic diatom analysis.

Main Methods:

• A dataset of 20,000 images of diatoms and background from lung tissue was created.
• Four Convolutional Neural Network (CNN) models (VGG16, ResNet50, InceptionV3, Inception-ResNet-V2) were trained and validated.
• Performance was evaluated using metrics like recall, precision, accuracy, and F1 score.

Main Results:

• InceptionV3 demonstrated superior performance with a recall rate of 89.80% and precision of 92.58%.
• VGG16 and Inception-ResNet-V2 showed comparable, acceptable performance.
• ResNet50 exhibited the lowest performance with a 55.35% recall rate.

Conclusions:

• InceptionV3 is the most effective model for diatom identification and feature extraction.
• The Inception-dependent model's targeted feature extraction makes it ideal for forensic diatom examination.