news

한어Русский языкEnglishFrançaisIndonesianSanskrit日本語DeutschPortuguêsΕλληνικάespañolItalianoSuomalainenLatina

Mosquitoes rank first among animals that cause death to humans.

According to statistics from the World Health Organization (WHO), about 725,000 people die from diseases caused by mosquito bites every year, and malaria alone caused 608,000 deaths in 2022. In contrast, other deadly animals such as snakes, dogs (by spreading rabies), crocodiles, etc., although they also cause a considerable number of deaths, are much smaller than mosquitoes.

Comparison of the number of people who die from various animals each year

Although the World Health Organization declared the complete eradication of malaria as early as 1955, in many areas, the eradication of malaria has become a game of "whack-a-mole." Every time malaria is considered to be under control, it will soon appear in another form, and it has not been effectively controlled to this day.

If we want to eliminate diseases transmitted by mosquitoes, we must first detect and distinguish different types of mosquitoes. This is not only because different types of mosquitoes carry different diseases, but also because different mosquitoes have different survival characteristics.

This means that if you can identify the types of mosquitoes, you can use their different characteristics to kill them. For example, for mosquitoes that live and forage outdoors, you can eliminate breeding grounds, and for mosquitoes that live indoors, you can use mosquito nets. This has proven to be an effective mosquito control measure in many areas.

Bill Gates, the former world's richest man, recently shared a new technology for this purpose - VectorCam. It can identify the mosquito species, gender, whether it sucks blood and lays eggs by just taking a photo of the mosquito through the application:

In the fight against mosquitoes, we finally saw our opponent clearly.

Bill Gates introduces the technology in a video

Using AI computer vision to "see" mosquitoes

According to VectorCam's official introduction, the system uses a new convolutional neural network VectorBrain to identify mosquito species, gender and abdominal state.

As an AI model specifically trained to identify mosquitoes, VectorBrain can accurately identify six major mosquito vectors, including major malaria vectors, with an accuracy rate of over 90% in resource-constrained environments.

VectorBrain is a multi-task EfficientNet architecture designed for mosquito classification, which outputs species, sex, and abdomen state simultaneously. The architecture consists of a feature extractor and a branch structure, with each branch corresponding to a classification task.

In terms of mosquito identification, VectorBrain uses a lightweight YOLO model that can locate mosquitoes in real time and use the detected coordinates to crop out images containing only the mosquitoes themselves for better identification.

The image illustrates the various stages of classifying a mosquito image. First, the full mosquito image to be classified is shown (a). Then, the mosquito image is cropped according to the coordinates using the YOLO algorithm and a series of image transformations are performed in preparation for classification (b). Finally, the output of the classification algorithm is shown, determining the mosquito species in the image (c).

Specifically, the accuracy, recall, and mean average precision (mAP) of the YOLO model during training and validation were 96.00%, 90.50%, and 95.87%, respectively. The accuracy of the category classification model was 92.40±2%, the accuracy of the gender classification model was 97.00±1%, and the accuracy of the abdominal state classification model was 83.20±3.1%.

(a) is the performance index of the YOLO model in training and verification, and (b) is the mosquito detection case of the model.

Confusion matrix and accuracy by species, gender, and abdominal status

In the paper provided by VectorCam, the YOLOv5 model it is using is compared with the Faster R-CNN model, which is widely used in various target detection tasks. YOLOv5 Small performs better in terms of the number of parameters, model size, mAP, and running time.

Barefoot doctors can also get started quickly

In addition to being a more targeted large model, VectorCam has also simplified its specific operations to better adapt to the actual conditions in malaria-transmitted areas.

Specifically, VectorCam consists of a set of specialized imaging equipment and a mobile phone application. The hardware components include a light box with a built-in 15x macro lens, a mobile phone case design, and a docking station. The hardware also includes Eppendorf tube holders and mosquito trays and perforated specimen ID sheets to better store these mosquitoes.

VectorCam's software is an Android-based application that is said to be able to identify more than 39 mosquito types, including common mosquitoes as well as some specific mosquito types that are more likely to carry diseases, and its algorithm has been optimized to run on lower-end Android phones.

VectorCam's mobile interface

The system's workflow involves placing collected mosquitoes into the hardware, capturing magnified images of the mosquitoes using a smartphone app, and storing the mosquitoes in uniquely labeled Eppendorf tubes for subsequent molecular verification.

The entire workflow of the VectorCam system

Imaging and loading tasks require only two users: one for imaging and the other for loading and storing mosquitoes. No entomological expertise is required, and even rural health teams can operate VectorCam with simple training.

In addition to being easy to use, VectorCam also has the advantage of being able to display the regional distribution of mosquitoes in a more intuitive way, allowing decision makers to better understand the situation and judge the local area based on the mosquito species, gender, abdominal condition, etc., and take appropriate measures to respond, thereby promoting the process of completely eradicating the stubborn disease of malaria.

The picture shows how the network "perceives" the characteristics of mosquitoes at different levels, and displays their species, abdominal state, and gender distribution in an intuitive way.

Using AI computer hearing to "recognize" mosquitoes

In terms of using mobile phones to detect mosquitoes, Bill Gates also introduced another achievement - HumBug.

The new system is a set of machine learning algorithms that can use the acoustic characteristics (sounds) of mosquito flight notes captured by smartphones to identify mosquito species by the sound of their wings beating. (It turns out that different mosquito species have different wing beat speeds, which can vary in sound, depending on their size, age, and ambient temperature.)

HumBug project specific workflow

And more importantly, Humbug does not require the use of special equipment to collect mosquitoes like VectorCam, which further simplifies the mosquito detection process.

It’s still early days for Humbug, but if successful it could enable more automated and continuous monitoring.

When introducing these technologies, Bill Gates also expressed some concerns, not about technical difficulties, but about other political and economic factors:

One of the biggest challenges we face is not scientific, but financial and political.