Researchers develop advanced machine learning to count blood cells

The use of machine learning to do blood cell counts for illness detection rather than costly and often less reliable cell analyzer devices has proven labor-intensive, as the training of the machine learning model requires a tremendous amount of manual annotation activity by humans. Scientists at Beihang University, on the other hand, have created a new training system that automates most of this process.

A study published in the journal of Cyborg and Bionic Systems on April 9^th, 2022, describes their new training plan.

The amount and type of cells in the blood play an important role in illness diagnosis, but the cell analysis procedures generally used to count blood cells—that involve the detection and evaluation of physical and chemical features of cells suspended in fluid—are costly and time-consuming.

Furthermore, due to many influences such as temperature, pH, voltage, and magnetic field that might mislead the equipment, the accuracy of cell analyzer devices is only approximately 90%.

Much recent research on alternatives has focused on the use of computer algorithms to do “segmentation” on images of blood obtained by a high-definition camera linked to a microscope to increase accuracy, keep things simple, and minimize costs. Segmentation entails algorithms that label what emerges in a photo pixel by pixel, in this example, which sections of the image are cells and which are not basically, counting the number of cells in an image.

Such techniques function well when dealing with photos with only a single type of cell, but they struggle when dealing with images with several types of cells. In recent times, scientists have drawn to convolutional neural networks (CNNs), a type of machine learning that mimics the network components of the human visual brain, to overcome the challenge.

To accomplish this, the CNN must first be “trained” to recognize what is and is not a cell on tens of thousands of images of cells that have been manually tagged. It recognizes and counts the cells in a novel, unlabeled image after that.

But such manual labeling is laborious and expensive, even when done with the assistance of experts, which defeats the purpose of an alternative that is supposed to be simpler and cheaper than cell analyzers.”

Guangdong Zhan, Study Co-Author and Professor, Mechanical Engineering and Automation, Beihang University

So Beihang Research scientists devised a novel training plan for CNNs, in this case, U-Net, a complete convolutional network segmentation model that has been widely utilized in medical picture segmentation since its inception in 2015.

The CNN is first trained on a batch of thousands of pictures with only one type of cell in the new training method (taken from the blood of mice).

These single-cell photos are automatically “preprocessed” by traditional algorithms that decrease noise, improve image clarity, and recognize the contours of objects in the image. They then do adaptive picture segmentation. This algorithm analyzes the different levels of grey in a black and white image, and if a portion of the image lies beyond a specified grey threshold, it is segmented as a separate item.

The procedure is adaptive because, rather than segmenting off parts of the picture segments according to a predetermined grey threshold, it does so based on the image’s local properties.

The model is fine-tuned using a limited handful of manually annotated photos of different cell kinds after the single-cell-type training set is shown to it. In comparison, some manual annotation is still required, but the number of photos that need to be tagged by humans has decreased from thousands to just 600.

The researchers employed a typical cell analyzer on the identical mouse blood samples to perform an unbiased cell count against which they could contrast their new training system. They discovered that their training strategy segmented multiple-cell-type images with an efficiency of 94.85%, which is the same as training with manually annotated multiple-cell-type images.

Further complex versions can use the method to address more sophisticated segmentation challenges.

The scientists intend to build a very automatic approach for annotating and training models in the future, as the new training method still requires some manual annotation.