Multi-View Pareto Optimization for Minimal-Diagnostic-Set Identification of Disease Vectors

Area of Science:

• Entomology
• Computer Science
• Public Health

Background:

• Accurate disease vector identification is vital for public health.
• Distinguishing morphologically similar species requires specialized taxonomic expertise and extensive data.
• Current methods can be resource-intensive and time-consuming.

Purpose of the Study:

• To introduce MVP-Net, an AI framework for efficient disease vector identification.
• To extract a minimal set of diagnostic anatomical views from multi-view imagery.
• To reduce computational costs and data requirements in vector surveillance.

Main Methods:

• Developed MVP-Net, an AI-driven framework utilizing multi-view imagery.
• Evaluated the framework on Calyptratae and Culicidae datasets from Shanghai surveillance.
• Employed Pareto-based view optimization to reduce input data while maintaining performance.

Main Results:

• MVP-Net achieved high Top-1 accuracies (87.04% for Calyptratae, 100% for Culicidae) with all views.
• Optimized view selection reduced input to 5 views (Calyptratae) and 2 views (Culicidae).
• Computational costs decreased by 37.49% (Calyptratae) and 81.82% (Culicidae) with comparable performance.

Conclusions:

• MVP-Net effectively reduces view redundancy in multi-view identification.
• The framework offers a practical approach to optimize regional vector surveillance workflows.
• AI-driven view optimization enhances efficiency and maintains identification accuracy for disease vectors.