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

Accurate identification of disease vectors is crucial for public health, yet distinguishing morphologically similar species demands significant taxonomic expertise and data resources. This study proposes MVP-Net, an AI-driven framework designed to extract a minimal sufficient set of diagnostic anatomical views from multi-view imagery for efficient identification. The framework was evaluated on regionally collected datasets of Calyptratae (8 views) and Culicidae (11 views) from routine surveillance in Shanghai. Under all-view fusion, MVP-Net achieved Top-1 accuracies of 87.04% for Calyptratae and 100% for Culicidae. After Pareto-based view optimization, the required input was reduced to 5 views for Calyptratae and 2 views for Culicidae, lowering computational cost by 37.49% and 81.82%, respectively, while retaining comparable classification performance (86.11% for the recommended Calyptratae configuration and 100% for the recommended Culicidae configuration). These results show that MVP-Net can reduce view redundancy while preserving comparable identification performance within the current Shanghai surveillance setting, providing a practical approach for optimizing regional multi-view auxiliary identification workflows.