Col-Ovo: Smartphone-based artificial intelligence for rapid counting of Aedes mosquito eggs under field conditions

Col-Ovo is a smartphone-based artificial intelligence tool that enables rapid, automated counting of *Aedes aegypti* eggs from field images with over 95% accuracy, effectively overcoming the manual processing bottleneck in large-scale mosquito surveillance programs.

The Gist

Imagine a world where tiny, invisible soldiers are spreading dangerous diseases like dengue, Zika, and yellow fever. These soldiers are mosquitoes, specifically the Aedes aegypti species. To stop them, health workers need to know exactly how many are around.

Traditionally, they use simple traps called ovitraps (egg traps). Female mosquitoes lay their eggs on little strips of cloth inside these traps. If you count the eggs, you know how many mosquitoes are in the neighborhood.

The Problem: The "Needle in a Haystack" Nightmare
Here's the catch: Mosquito eggs are microscopic. They are smaller than a grain of sand. To count them, a human expert has to look at a dirty, stained piece of cloth under a magnifying glass.

• It takes 15 minutes to count just one trap.
• The cloth is often covered in mud, yeast, and molasses (used to attract the mosquitoes), making the eggs hard to see.
• If you have 1,000 traps, you need a massive team and weeks of time. It's like trying to count every single grain of sand on a beach by hand.

The Solution: Col-Ovo (The "Super-Scanner")
The authors of this paper, Juan and Diego, built a digital superhero named Col-Ovo. It's an Artificial Intelligence (AI) tool that lives on your smartphone.

Think of Col-Ovo as a super-powered pair of glasses that never gets tired.

1. Snap a Photo: A health worker takes a picture of the dirty cloth with their phone.
2. Send it: They send the photo via WhatsApp (even if the image gets a little blurry from compression, Col-Ovo doesn't mind).
3. The Magic: Col-Ovo instantly scans the photo, finds every single tiny egg, and counts them in 25 seconds.

How Did They Train It?
You can't just teach an AI to count eggs; you have to show it thousands of examples.

• The Classroom: They created a digital classroom called OviLab.
• The Students: They fed the AI 275 real photos of dirty, muddy, yeast-stained cloths from the field.
• The Teacher: A human expert manually counted the eggs in these photos to create the "answer key."
• The Result: The AI learned to spot eggs even when they were hiding under dirt or stuck together in clumps.

Why is This a Game-Changer?

• Speed: It's 11 times faster than a human. What used to take 15 minutes now takes less than half a minute.
• Accuracy: It counts just as well as the best human experts, but without getting tired or making mistakes because of eye strain.
• Accessibility: You don't need a fancy microscope or a supercomputer. You just need a regular smartphone and an internet connection. It works in remote villages just as well as in big cities.

The Big Picture
Imagine a fire alarm system. If you wait until the fire is huge to call for help, it's too late. By using Col-Ovo, health workers can spot a "mosquito fire" (an outbreak) the moment it starts. Because the counting is so fast, they can check thousands of traps in a single day, map out exactly where the mosquitoes are, and stop the disease before it spreads.

In short, Col-Ovo turns a slow, painful, manual chore into a quick, digital snap, giving communities a powerful new weapon to fight mosquito-borne diseases.

Technical Summary

1. Problem Statement

Vector-borne diseases such as dengue, Zika, chikungunya, and urban yellow fever are transmitted by Aedes aegypti and Aedes albopictus mosquitoes. Surveillance programs increasingly rely on ovitraps (traps that attract female mosquitoes to lay eggs) to monitor population density and guide control strategies. A specific low-cost strategy, OviCol, uses handmade traps with bioattractants (molasses and yeast) and hot water to inactivate eggs.

However, a critical operational bottleneck exists: manual egg counting.

• Scale: Large-scale surveillance generates thousands of egg substrates (strips).
• Complexity: Aedes eggs are microscopic (~0.065 mm²) and often clustered.
• Conditions: Substrates are frequently stained by bioattractants, covered in sand/debris, and photographed under variable lighting.
• Limitations: Manual counting is slow (~15 minutes per sample), labor-intensive, prone to human error/fatigue, and difficult to scale. Existing AI solutions often require laboratory-grade images, specialized hardware, or complex preprocessing, making them unsuitable for routine field deployment.

2. Methodology

Data Collection and Curation

• Samples: 275 oviposition substrates (white nonwoven napkins, 20.5 cm strips) collected from wild Ae. aegypti populations in Santa Marta, Colombia.
• Conditions: Samples included heavy staining from molasses/yeast, sand, and particulate debris to simulate realistic field conditions.
• Image Acquisition: Images were captured using an iPhone X (12 MP). Crucially, images were shared via WhatsApp to simulate real-world compression artifacts and JPEG noise.
• Ground Truth: A dedicated web platform, OviLab, was developed for expert annotation. An entomology expert manually annotated 27,562 eggs across the dataset. The platform supports up to 200% magnification to distinguish clustered eggs.
• Dataset Split: 70% training, 15% validation, 15% testing (split at the image level to prevent data leakage).

Preprocessing and Tiling Strategy

• Tiling: Due to high-resolution images (up to 4032 pixels) and tiny egg sizes (~10 pixels), images were divided into 640 × 640 pixel tiles with a 20% overlap (stride of 512).
• Augmentation: Intensive data augmentation was applied to simulate field variability, including 360° rotation, scaling, flips, mosaic augmentation, mixup, copy-paste, random erasing, and HSV color adjustments.
• Negative Sampling: Negative tiles (no eggs) were included at a 1:4 ratio to prevent background bias.

Model Architecture and Training

• Algorithm: YOLOv11m (You Only Look Once version 11, medium model).
• Training Configuration:
  • Optimizer: Stochastic Gradient Descent (SGD).
  • Hyperparameters: Initial learning rate 0.01, batch size 8, 150 epochs.
  • Small Object Optimization: The bounding box loss weight was increased from 7.5 to 12.0 to improve detection of tiny eggs.
  • Inference: Non-Maximum Suppression (NMS) with a confidence threshold of 0.25 and IoU of 0.3 to merge detections across tile boundaries.
• Infrastructure: Trained on an NVIDIA RTX 5090 GPU (32 GB VRAM) in ~36 minutes.

Validation

• Comparison: An independent subset of 30 samples was evaluated using three methods:
  1. OviLab: Reference standard (virtual manual counting with high magnification).
  2. Human Expert: Direct visual counting.
  3. Col-Ovo: The AI system.
• Difficulty Stratification: Samples were classified into Easy, Medium, and Hard (high contamination/density).
• Statistical Tests: Friedman test, Wilcoxon signed-rank, Spearman correlation, Bland-Altman analysis, and Lin's Concordance Correlation Coefficient ($\rho_c$).

3. Key Contributions

1. Col-Ovo System: A lightweight, web-based AI tool for automated Aedes egg counting that operates directly on smartphone images without preprocessing.
2. Field-Ready Robustness: The model is trained specifically on compressed, stained, and debris-covered images, tolerating artifacts from messaging apps like WhatsApp.
3. OviLab Platform: An open-access digital infrastructure for annotation, curation, and dataset management, enabling future retraining and expansion to other Aedes species.
4. Operational Efficiency: A workflow that reduces processing time from minutes to seconds, enabling real-time surveillance.
5. Scalability: The system requires no specialized hardware, making it accessible for community-level vector control programs in resource-limited settings.

4. Results

Performance Metrics

• Detection Accuracy: Achieved a mAP@0.5 of 0.825 and mAP@0.5:0.95 of 0.332.
• Precision/Recall: Precision 0.804, Recall 0.794, F1-score 0.80.
• Agreement with Reference:
  • Col-Ovo vs. OviLab: 88% agreement; statistically indistinguishable ($p=0.252$).
  • Human vs. OviLab: 95% agreement; statistically significant difference ($p=0.009$), indicating human underestimation.
• Concordance ($\rho_c$):
  • Col-Ovo maintained high concordance across all difficulty levels (Easy: 0.993, Medium: 0.924, Hard: 0.952).
  • Human counting dropped significantly in "Hard" samples ($\rho_c = 0.794$).

Accuracy and Bias

• Regression Analysis: Col-Ovo showed a slope of 0.785 (closer to ideal 1.0) compared to Human counting (0.561), indicating better proportional agreement at high egg densities.
• Error Rates: Col-Ovo had a lower Mean Absolute Error (23.4 eggs) compared to humans (28.6 eggs).
• Bland-Altman: Col-Ovo exhibited narrower limits of agreement ($-119.1$ to $+89.5$) compared to humans ($-220.2$ to $+167.2$).

Efficiency

• Processing Time: Col-Ovo averaged 21 seconds per sample (SD 9.5s).
• Comparison: This is an 11.3-fold speed increase over the OviLab reference method (242 seconds) and manual counting (~15 mins).
• Throughput: 3.2 samples/minute vs. 0.25 samples/minute for manual methods.

5. Significance

• Operational Impact: Col-Ovo removes the primary bottleneck in ovitrap-based surveillance, allowing health programs to process large networks of traps rapidly.
• Public Health: Faster data turnaround enables earlier detection of vector density spikes, facilitating timely interventions against arboviral outbreaks.
• Accessibility: By functioning on standard smartphones and handling compressed images, the tool democratizes advanced surveillance capabilities, removing the need for expensive scanners or high-end workstations.
• Scalability: The integration of OviLab and Col-Ovo creates a feedback loop where new data can be used to retrain the model, adapting it to new ecological contexts or different mosquito species.
• Community Engagement: The speed and ease of use support community-led control strategies, empowering local teams to monitor and manage vector populations effectively.

In conclusion, Col-Ovo represents a significant advancement in vector surveillance technology, successfully translating AI from a research prototype into a practical, field-deployable tool that balances high accuracy with extreme operational efficiency.