Our Work - Puerto Rico Science, Technology & Research Trust

Deep Learning Models Aid Research to Prevent Mosquito-Borne Illnesses.

 

OVERVIEW

The world’s deadliest animal is not a shark or a snake, it is the mosquito. The tiny creatures have killed 32 times as many people as every war in human history combined. The Puerto Rico Science, Technology & Research Trust and its Vector Control Unit (PRVCU) has been working to help prevent and manage diseases spread by mosquitoes, to gain an understanding of why many mosquitoes have become immune to insecticides approved by the FDA.

Researchers have spread out across the island, capturing different mosquito species in traps; monitoring and testing them for viral presence and insecticide resistance; and labelling and classifying them.  Manually capturing and classifying thousands of mosquitoes in different locales can take many months before any specific patterns can emerge.

Wovenware partnered with the PRVCU to develop an RPA solution composed of deep learning models and other processes to automate the identification and classification of mosquito species. It significantly reduced manual work, allowing entomologists to spend more time in more advanced research activities.

PRSTRT Picture

Roles

  • Project Leadership
  • Data Engineers
  • Data Scientists

Deliverables

  • Exploratory Analysis
  • Training Dataset
  • Classification Model
  • Results Documentation

THE CHALLENGES

What We Faced

  • Classifying very small objects in an image.
  • Correctly classifying a large number of mosquitoes that are mutilated by the weather or traps.
  • Experimenting with new architectures that could change the technical approach to the problem.
  • Designing a human + machine RPA-based workflow to integrate the model with day-to-day research operations conducted by scientists.

OUR APPROACH

How We Helped

Wovenware proposed an RPA solution to automate counting the total number of mosquitoes in traps, classifying the species and the gender. The team:

  • Built a Convolutional Neural Network (CNN) to count total number of mosquitos
  • Built two CNNs to identify mosquito species and gender
  • Trained the models with a dataset of 5,000 images

Wovenware trained models on our state-of-the-art local 8 GPU cluster server optimized for deep learning training.

Technologies

  • RETINANET

    RETINANET

  • RESNET

    RESNET

  • TensorFlow

    Tensorflow

RESULTS

Client Success

The automated classification process was four times faster than the manual process and it freed up time for the entomologists to focus on analytics and research to understand why many mosquitoes have become immune to insecticides approved by the FDA. The team is experimenting with camera setup to automate the tasks of taking pictures from traps. Future improvements can be made by including time-series analysis, genome analysis, and remote monitoring and classification.