Spectral intelligent detection for aflatoxin B1 via contrastive learning based on Siamese network

Area of Science:

• Agricultural Science
• Food Safety
• Machine Learning

Background:

• Aflatoxins are toxic metabolites produced by fungi, commonly found in staple foods like peanuts and corn.
• Exposure to aflatoxins poses significant health risks, including liver damage and cancer.
• Accurate and rapid detection of aflatoxins is crucial for food safety and public health.

Purpose of the Study:

• To develop an innovative model for precise aflatoxin detection in peanuts and corn.
• To enhance feature discrimination for improved detection accuracy.
• To establish a method for continuous, real-time monitoring of aflatoxin contamination.

Main Methods:

• The study introduces the MSGhostDNN model, integrating contrastive learning with multi-scale convolutional networks.
• Grad-CAM was applied to identify key wavelengths for detection, focusing on 416 nm and 40 optimal wavelengths.
• A task dimensionality reduction approach was employed for continuous learning and spectrum monitoring.

Main Results:

• The MSGhostDNN model achieved a detection accuracy of 97.87% with a pre-trained model.
• Refinement using Grad-CAM and focusing on 40 key wavelengths resulted in 97.46% accuracy.
• The dimensionality reduction approach enabled effective ongoing aflatoxin spectrum monitoring.

Conclusions:

• The MSGhostDNN model offers a highly accurate and efficient method for aflatoxin detection in agricultural products.
• The identified key wavelengths and dimensionality reduction approach facilitate rapid online monitoring.
• This research provides a precedent for the online detection of similar foodborne toxins, enhancing food safety protocols.