YOLito: A generalizable model for automated mosquito detection.

Area of Science:

• Vector biology
• Computational biology
• Ecological research

Background:

• Mosquito behavior research is crucial for ecology, evolution, and disease control.
• Current methods for observing mosquito behavior are labor-intensive, limiting scalability and reproducibility.

Purpose of the Study:

• To develop an automated, AI-driven system for mosquito detection and behavioral quantification.
• To enhance the throughput and reproducibility of mosquito behavioral assays.

Main Methods:

• Developed YOLito, a domain-generalized AI model using the Ultralytics YOLO framework and Slicing-Aided Hyper Inference (SAHI).
• Trained YOLito on a diverse dataset of 38,547 annotated images from 35 experimental setups, covering multiple mosquito species and assay types.
• Validated YOLito's performance on unseen data, achieving high precision (0.95) and recall (0.91).

Main Results:

• YOLito accurately detects and quantifies mosquito behavior across varied backgrounds and imaging conditions.
• The model demonstrates generalization across mosquito species (Aedes, Anopheles, Culex) and experimental setups (blood-feeding, sugar-feeding, oviposition).
• Achieved high performance metrics: precision = 0.95, recall = 0.91 on unseen data.

Conclusions:

• YOLito transforms traditional mosquito behavioral assays into scalable and reproducible platforms.
• The open-source toolkit facilitates high-throughput extraction of behavioral metrics, standardizing research.
• This AI framework bridges computer vision advancements with vector biology applications.