Use case

Detecting worms in microscope images

Scanning through microscopic images to find, identify and count worms. Sounds interesting? We thought so too.
Detecting worms in microscope images

Executive summary

Some treasures can only be found by experts, such as doctors or veterinary specialists. Or might it be possible to teach a machine to do so? Our customer wanted to find out: They needed a fast and reliable way to detect different kinds of worms in microscope images. After initial training, they are now getting a response within a matter of seconds from colabel and can provide much faster service to pet lovers.

Our customer

The company is an online veterinary, offering a wide range of products and services related to the health of pets and livestock. Their aim is to provide care and service for preventive animal care in a modern and digitized way. To compete with established veterinaries, they are in constant need to develop innovative products and services: fast, convenient and affordable – at highest quality standards.

Challenge

Checking pets for worms used to be a big burden for pet owners. Every few months, a veterinary has to be consulted in order to ensure the animal's well-being. Our customer has developed a service that enhances this experience from a time and convenience standpoint: as opposed to visiting the veterinary, the pet owner orders a set to take stool samples at home and return them via mail for laboratory analysis.

In order to recommend the right treatment, two factors are relevant:

  1. Type of parasite
  2. Number of parasites

Traditionally, the samples were analyzed by veterinary specialists. However, as the company started selling more and more units, this demand could no longer be covered in-house but had to be done by contract veterinaries. As a result, the process became more complex for our customer to manage and the time between receipt of the sample and result became larger.

Objective

The company was looking for alternative ways to work with the samples they received. Creating a microscopic image could be done in-house in a very efficient manner. However, the analysis still had to be done by experts, even though the range of possible outcomes is very limited. Our customer therefore sought out for a solution that was not only efficient but also very accurate: the result was used not only to inform the pet owner but also recommend appropriate treatment and medicine.

What our customer built

With that objective in mind, our customer decided to go for a process with two training stages.

  • Stage 1: Analyze the image on content to find out if there are any worms in the sample.
  • Stage 2: Run the positive images through an object detection algorithm in order to count how many worms there are.

Why make it more complicated? Sometimes it helps to filter out some of the noise in the data. Neural networks show astonishing performance out of the box on a number of tasks but in this specific case, it required a narrower target for the actual identification and counting. This was one case where the easiest way was not the best.

Example for Stage 2: The object detection and counting using colabel.

In order to get the same performance in a production setting, these two stages would have to be carried over to the live process. Our customer decided to initially connect the process via Google Sheets and Zapier.

Result

This configuration returns a response within a matter of seconds after the microscope images had been sent to our system. Mapped with different treatment recommendations, the process could be automated to a large degree, leading to a high perceived responsiveness by their own customers – and quick help for their pets.

Our customer was able to build a key element of their service in an automated way. Through using colabel, a task thus far only carried out by experts could be digitized in a way that would otherwise require serious investments in deep learning on the engineering side – a function that simply does not exist in their company and possibly never will. This still does not make the whole project simple, but it makes it possible.

Ready to build your own?