91��ý

In modern medicine, time is sometimes measured in minutes. This is especially evident in cases of acute stroke — one of the most dangerous neurological conditions, where every lost hour can cost a patient their life or lead to irreversible consequences. That is why the concept of the “golden hour” exists in medicine — a short but critical period during which timely diagnosis and treatment can dramatically change the outcome of the disease.

However, there are still significant challenges in achieving fast and accurate stroke diagnosis. A shortage of specialists in neuroimaging, particularly in remote regions, the time required to interpret CT and MRI scans, and discrepancies between expert evaluations all complicate the diagnostic process. In addition, modern scanners generate enormous volumes of data that are difficult to analyze manually within limited timeframes.

In this context, artificial intelligence technologies are becoming increasingly important. In recent years, deep learning methods have demonstrated remarkable effectiveness in analyzing medical images. However, most of these systems function as “black boxes”: they provide results without explaining how decisions are made. In medicine, where transparency and trust are essential, this remains a major limitation.

To address this issue, researchers at the al-Farabi Kazakh National University are implementing a large-scale research project entitled “Development of Combined Deep Neural Network Models for Interpretable Analysis of Medical Images.” The project, running from 2025 to 2027, brings together experts in artificial intelligence, machine learning, image processing, and clinical radiology.

The project’s main objective is to create a model capable not only of accurately diagnosing stroke, but also of explaining its decisions. The proposed system combines convolutional neural networks with a Soft Decision Tree architecture. This approach enables the system not only to identify the presence of stroke, but also to demonstrate which image features led to its conclusion.

The practical component of the project focuses on developing a system for detecting acute stroke using CT and MRI brain scans. To train the model, researchers are creating a database of annotated medical images in which specialists manually identify pathological changes. The project also employs weak supervision and self-learning methods, allowing the system to work effectively even with partially labeled data.

One of the promising research directions involves federated learning, which allows models to be trained on data from multiple medical institutions without transferring the data itself, thereby preserving patient confidentiality. In addition, the project explores interpretable architectures based on Kolmogorov–Arnold Networks.

Researchers are also developing a user-friendly interface that will allow medical professionals to upload scans, receive analysis results, and visualize the features underlying the system’s decisions.

At the final stage, the system will undergo clinical validation, and its performance will be compared with the assessments of experienced radiologists. The goal is not to replace physicians, but to provide them with an intelligent tool capable of improving the speed and accuracy of clinical decision-making.

The KazNU researchers’ project is not merely a technological development, but an important step toward creating transparent and trustworthy artificial intelligence in medicine, where confidence in technology is essential for its practical value.