BelSU Scientists Develop Innovative Program for Recognizing and Classifying Leukocytes in Blood Cell Images

Led by Associate Professor Vladimir Mikhelev from the Department of Mathematical and Software Information Systems at Belgorod National Research University (BelSU), a dedicated team from the Institute of Information and Digital Technologies has created an advanced system capable of automatically identifying various types of white blood cells, or leukocytes, in microscopic images.

In clinical diagnostics, speed is of the essence, and this new program is designed to meet that critical need. While automation in blood cell analysis has been around for some time, existing methods often fall short – either being prohibitively expensive or lacking the versatility required for diverse laboratory settings. Given that each lab produces images with unique characteristics – varying in brightness, quality, and colour – accurate recognition has long been a challenge. The researchers aimed to develop a universal solution that could effectively analyse images from different laboratories and microscopes.

Their innovative approach combines two powerful methodologies: expert rules and a trained neural network. The first component relies on established expert rules that encapsulate the knowledge of specialists involved in manual analysis. These rules assess key characteristics of leukocytes, such as shape, colour, surface structure, and the distribution of colours and shapes within the cells. By employing Bayesian inference, the program not only leverages existing knowledge but also incorporates new data gathered during observations, further refining its accuracy.

While systems based on expert rules offer transparency and explainability, they can be subjective – different specialists may describe the same colour in various ways. To address this, the second component of the program utilizes a deep learning neural network. This AI-driven system autonomously detects blood cells in images and classifies them into one of five categories: neutrophils, lymphocytes, monocytes, eosinophils, and basophils. With proper configuration and extensive training data, these neural networks can achieve impressive results. However, they are not infallible; understanding the rationale behind their decisions is crucial in medical contexts, yet traditional neural networks often lack this clarity.

The BelSU team has successfully merged the strengths of both expert rules and neural networks. Their program processes each image through these two methods, combining the outcomes for a final classification. If both methods agree on a result, it is deemed conclusive. In cases of disagreement, the automated analysis is forwarded to a specialist for manual verification.

Remarkably, this innovative program processes images in under a second while achieving an impressive accuracy rate of approximately 93% k – comparable to many modern complex systems, including large neural networks. What sets this development apart is its commitment to transparency; the program meticulously records recognized features and activated rules, enabling experts to trace the decision-making process and make necessary adjustments.

“The development of mathematical models and hybrid methods that integrate the strengths of neural networks with interpretable fuzzy models and Bayesian approaches is vital for medical diagnostics,” said Mikhelev. “From a practical standpoint, they ensure high accuracy, transparency, and ease of expert analysis while allowing for necessary adjustments in automated processes.”

After conducting computational experiments on real medical data, the researchers validated the effectiveness of their hybrid method for classifying leukocytes. This innovative software package is poised to be implemented in medical and laboratory information systems. Looking ahead, the researchers plan to refine the model further and enhance its adaptability and classification accuracy through active learning techniques.