91��ý

A research team at 91��ý Kazakh National University (KazNU) is implementing an ambitious scientific project aimed at creating a new generation of intelligent systems for medical image analysis. The project, titled "Development of Combined Deep Neural Network Models for Interpretable Medical Image Analysis," spans 2025 to 2027 and is supported by grant funding exceeding 119 million tenge under the national priority direction of Advanced Manufacturing, Digital and Space Technologies.

Acute stroke is among the most time-critical emergencies in medicine. The concept of the "golden hour" is well established: rapid and accurate diagnosis in the first hours following a cerebrovascular event can dramatically alter patient outcomes — reducing mortality and limiting lasting neurological damage. Every minute of delayed diagnosis can mean the difference between recovery and permanent disability.

Despite significant technological advances, neuroimaging today faces persistent systemic challenges. Access to advanced diagnostic tools remains limited across many regions. There is a critical shortage of qualified specialists capable of promptly and accurately interpreting MRI and CT scan data. Inter-expert variability in assessments introduces inconsistency into clinical decision-making. These factors together create an urgent need for intelligent systems that can rapidly and reliably analyze large volumes of heterogeneous data — even when annotated training examples are scarce.

Artificial intelligence and deep learning have already demonstrated remarkable performance in medical image processing. Yet this is precisely where the central challenge addressed by the KazNU team becomes most acute: the majority of high-performing neural network models operate as black boxes. They produce a diagnosis but offer no explanation for it. In clinical practice, this is unacceptable. A physician cannot — and should not — blindly trust a system whose reasoning they cannot follow or verify.

The defining scientific novelty of this project lies in a fundamentally different approach to model design. Rather than adding interpretability as an afterthought — through supplementary analysis of an already-trained network — the KazNU researchers embed transparency directly into the neural network architecture itself.

The proposed model combines a convolutional neural network (CNN) with a Similarity-Based Decision Tree (SBDT) classifier. The hidden features extracted in the final convolutional layer of the network are transformed into semantically meaningful concepts — object parts, anatomical landmarks, pathological changes — that specialists can interpret in their own domain language. The decision tree then visualizes the recognition logic as a sequence of rules that are legible to a clinician without requiring expertise in machine learning.

This approach — termed "a priori interpretation" — distinguishes the project from the overwhelming majority of existing systems. Transparency is built into the model from the outset, not derived from post hoc analysis. The model is designed to make its internal reasoning visible and understandable, providing physicians with a hierarchically organized explanation of each classification decision (see Figure 1 and Figure 2 in the project documentation). The immediate goal is not merely to maximize predictive accuracy, but to create a system that a clinician can genuinely trust and engage with.

The project additionally plans to integrate Federated Learning (FL) and Kolmogorov–Arnold Networks (KAN). Federated Learning enables model training across multiple medical institutions without centralizing sensitive patient data, addressing one of the most significant practical barriers to deploying AI in healthcare — data privacy and regulatory compliance. The KAN architecture, grounded in the Kolmogorov–Arnold representation theorem, models complex multi-dimensional data in a more interpretable fashion by expressing any continuous function as a sum of continuous univariate functions, thereby enhancing the transparency of neural network decisions.

The project is structured around seven interconnected tasks, each contributing to the overarching goal of deployable, interpretable AI for clinical stroke diagnosis.

The first task involves designing and implementing a deep neural network architecture that unifies a fully convolutional network, a trainable visual pattern layer (analogous to a bag of visual words), and a Soft Decision Tree (SDT). This task also includes formalizing the requirements for how model outputs should be displayed to maximize comprehension by clinical users.

The second task concerns building a repository of annotated brain MRI scans. Medical experts — including the project's in-house radiologist — will select and annotate confirmed acute stroke cases alongside healthy controls, creating training and test datasets with expert-marked pathological regions. This dataset will form the empirical backbone of the project's machine learning pipeline.

The third task focuses on developing interpretable deep learning models that extract significant features from imaging data and explain their decisions through logical probabilistic statements that are accessible to end users. The resulting system should be able to present not just a classification result, but a chain of reasoning that clinicians can evaluate and challenge.

The fourth task addresses a persistent challenge in medical AI: the scarcity of fully and correctly annotated data. Neural networks employing weakly supervised and self-supervised learning approaches will be developed to work effectively with datasets containing incomplete or noisy annotations — a condition that accurately reflects the realities of clinical data collection.

The fifth task involves building a user-friendly interface for data input and result visualization, leveraging professional medical terminology to facilitate seamless interaction between the system and its intended clinical users.

The sixth task is the practical clinical validation of the developed models. The system will be deployed as a prototype intelligent tool for detecting acute stroke from MRI and CT brain scans, and its performance will be evaluated in a real clinical environment.

The seventh task involves a rigorous retrospective comparative analysis: the system's sensitivity, specificity, positive predictive value, and negative predictive value will be benchmarked against the assessments of expert radiologists. This analysis will establish the real-world diagnostic utility of the tool and provide the evidence base for broader clinical adoption.

The primary training dataset is the RSNA Intracranial Hemorrhage Detection Challenge dataset, comprising 41,071 annotated brain CT images spanning five subtypes of intracranial hemorrhage: epidural, intraparenchymal, intraventricular, subarachnoid, and subdural. Experiments are designed to evaluate and interpret model outputs, identify key diagnostic features, and validate results for clinical applicability. Data quality assurance protocols include automated anomaly detection pipelines and manual validation of critical records by domain experts.

The project's outputs will be published in peer-reviewed international journals indexed in Scopus and Web of Science, contributing to the global body of knowledge on interpretable AI in medicine. The developed models and algorithms can serve as pretrained foundations for related medical diagnostic challenges, extending their impact well beyond stroke detection.

The project aligns with national priorities for the digitalization of healthcare and the development of domestic technological capabilities. By creating a reliable, explainable AI assistant for physicians, the KazNU team is taking a concrete step toward a future in which intelligent systems and clinical expertise work together — not in competition — to deliver faster, safer, and more accurate diagnoses for patients across Kazakhstan and beyond.