LOCAL CONSISTENCY METHOD FOR CREATING EXPLAINED MEDICAL IMAGE CLASSIFICATION MODELS

Abstract

An effective implementation of artificial intelligence systems in medical diagnostics is possible only with ensuring transparency and interpretability of decision-making processes, which is critically important for clinical acceptance and regulatory compliance. The specific challenge of medical AI applications is their "black box" nature that prevents understanding of diagnostic reasoning, leading to limited trust among clinicians despite achieving high accuracy rates. This paper presents the results of research on a novel local consistency method for creating interpretable AI systems in neuro-oncological diagnosis, which combines high-precision VGG-16 convolutional neural network with Decision Rules Network for generating clinically relevant explanations. The proposed hybrid architecture achieved diagnostic accuracy of 95.2% while maintaining the ability to generate logical rules based on IBSI-standardized features for four brain pathology types (glioma, meningioma, pituitary tumor, no tumor). Experimental validation demonstrated achievement of 89.3% local consistency between decisions of the primary and interpretable models, confirming reliability of personalized explanations for specific clinical cases. The developed approach addresses the fundamental trade-off between accuracy and interpretability by ensuring consistency only for local diagnostic instances rather than the entire data space, enabling personalized explanations adapted to specific clinical scenarios. The data obtained can be used in justification of trustworthy AI deployment in medical practice, as well as in development of more transparent and clinically acceptable diagnostic systems that combine technological efficiency with necessary transparency for healthcare applications.