Computing Artificial Intelligence for Science

Madeleine Clark

4 September 2023
Read for 5 minutes

Important points

  • Artificial intelligence (AI) is changing the slow, labor-intensive and expensive process of measuring small things in our science.
  • We are using AI to measure microscopic cells of hairs and harmful algae on cotton leaves.
  • AI can bring greater precision, speed and scale to tasks that are difficult for our scientists.

Identifying harmful algal cells in microscopic images is one way we’re using artificial intelligence to change the course of science.

There has been a lot of talk about how AI will change the way we work.

Whether you see AI as friend or foe in the workforce, there’s no doubt that this technology will change some jobs as we know them. This includes jobs in the world of science. Our researchers are exploring and responsibly embracing the use of AI in our work.

AI is a particularly useful tool for overcoming vexing biological limitations. For example, the human eye cannot detect objects smaller than about 0.2 mm. Traditionally, we use microscopes to overcome this. But once an object or substance is magnified, a highly trained expert must look at it to identify and measure objects of interest. Measuring samples or objects for science can be a slow, labor-intensive, and expensive process.

Here are some examples where we’ve leveraged the impressive advances in machine learning over the past decade to measure AI. These projects are advanced through our Machine Learning and Artificial Intelligence Future Science Platform, which brings together collaborators across the organization to advance machine learning for scientific discovery.

Counting hairs on cotton leaves

Dr Vivian Rowland and Moshiur Farazi observing cotton in a glasshouse in Canberra.

Working with our agricultural experts, we developed a model that can count the number of hairs on the back of a cotton leaf.

Leaf hair affects insect resistance, fiber yield and value of new cotton cultivars. Traditionally, experts in commercial breeding programs have looked at this and assigned a score between one and five.

First we developed AI models that can produce hairstyle scores similar to humans with 95 percent accuracy.

Dr. Moshiur Farazi is one of our experts in computer vision, which focuses on enabling computers to recognize and understand objects in images and video. He said HairNet2 is now going beyond automated methods.

“Training models to reproduce human ratings of hairiness can increase the speed and scale of analysis. However, these models reproduce the variability in the human estimates they are trained on,” Moshiur said.

“In HairNet2 we have created a new method of AI-enhanced scoring that is more robust, reliable and accurate.

“This model estimates the leaf area covered by hair by finding all the hairs on the leaf, which is not completely impossible for a human to do, but incredibly difficult and time-consuming.

“HairNet 2 was trained using about 1000 images where humans annotated each hair. This painstaking annotation process helped create an AI tool that can automate hair scoring beyond human comprehension,” he said.

New models are being deployed on the web interface for testing by breeders in the next cotton season. You can try the demo yourself quickly.

HairNet2 uses artificial intelligence to detect individual hairs on cotton leaves

Enumeration of microalgae cells

Harmful algal blooms are large populations of algae that can be toxic to both humans and animals.

To identify harmful algal blooms, extensive testing is done using expert microscopes and a counting chamber (a slide with precise gridlines that allow scientists to estimate the number of harmful algal cells in a liquid sample).

Dr Chris Jackett is an expert in object detection. As a postdoctoral research fellow at our National Research Collections Australia, he began working to augment these manual processes with AI.

“It’s a very time-consuming and labor-intensive exercise, and humans are limited in the number of samples they can process,” Chris said.

“Extended sessions at the microscope can also lead to health problems such as vision problems, poor posture, physical stress and headaches.”

In response to this challenge, we are training machine learning models to automatically detect harmful algae in images.

Harmful algal cells are detected by machine learning

Our team is systematically photographing and annotating algal strain samples from the Australian National Algae Culture Collection. We have also started using a range of AI tools to speed up the annotation process.

With this combined human/computer effort, we have created an annotated dataset for 15 different algal species to date, which is now being used to train AI models. Initial testing indicates that these models can successfully detect target strains with a high level of accuracy.

Using AI to provide faster and more accurate detection of toxic algae could have significant economic, environmental, and social impacts.

“Improving the speed and accuracy of detecting harmful algae can provide water managers with early warning signals indicating when and where blooms may occur,” Chris said.

AI-enhanced risk management and decision-making can help protect the health of the environment as well as coastal communities, consumers and Australian fisheries and aquaculture businesses.

How you can make AI count for your business

Many organizations are currently grappling with the potential of AI to transform their processes and businesses.

If your core business involves counting or recognizing objects, AI can be a friend.

Moshiur said the barrier to entry for those looking to apply AI to object detection is low.

“Five or 10 years ago you needed to train the model yourself, and to experiment with AI you needed a lot of data and computing power,” Moshiur said.

“If you have a very small amount of data, you can fine-tune open-source models with even a couple of hundred images to work on your problem.”

However, he said successful application of AI depends on asking the right questions.

“At the end of the day, most users want a black box where they can click a button and get the answer they want. But we need to unpack what they want the buttons to do and structure the data in such a way that those buttons can reach the correct answer,” Moshiur said.

“The best place to start is to sit down and explore the questions you need answers to, and consider the problems you can’t solve with human-driven methods or processes.”

Leave a Comment