Цифрова колориметрія та глибоке машинне навчання у діагностиці пошкоджень біотканин

• Datum: 2026
• Hauptverfasser: Кватернюк, С.М., Кватернюк, О.Є.
• Format: Artikel
• Sprache: Ukrainisch
• Veröffentlicht: Vinnytsia National Technical University 2026
• Schlagworte: судова медицина; мультиспектральна візуалізація; фізико-інформовані нейронні мережі; нейронечіткі системи; глибоке навчання; колірний простір; кінетика деградації гемоглобіну
• Online Zugang: https://oeipt.vntu.edu.ua/index.php/oeipt/article/view/868
• Назва журналу: Optoelectronic Information-Power Technologies

This paper investigates a critical challenge in forensic medical examination and clinical dermatology: the automated and objective estimation of the age of superficial biological tissue injuries based on digital and multispectral imaging. Traditional empirical approaches, grounded in the visual assessment of color dynamics in damaged skin areas, exhibit critically low precision and an excessive dependence on the expert's subjective experience, ambient lighting conditions, and the patient's physiological characteristics. The study substantiates and develops a comprehensive innovative paradigm that conceptually integrates the mathematical apparatus of digital colorimetry within the perceptually uniform CIELAB space with advanced deep machine learning architectures. A mathematical model of a neuro-fuzzy classification system is developed in detail to effectively simulate the spatio-temporal evolution of trauma, accounting for the inherent fuzziness of boundaries between the stages of blood degradation. Particular emphasis is placed on the implementation of Physics-Informed Neural Networks, which enable the integration of a system of partial differential equations – describing the biochemical kinetics of hemoglobin decay and bilirubin diffusion – directly into the multi-component loss function of the neural network. A profound analysis is conducted on the utilization of state-of-the-art semantic segmentation architectures for extracting complex spectral-spatial features. Furthermore, an optimized hardware-software pipeline is presented for the deployment and iterative training of models using Tensor Processing Units (TPUs) in a cloud environment. Based on hyperspectral analysis, it is demonstrated that the transition from heuristic approaches to end-to-end regression of injury age, considering physical constraints, significantly reduces the Mean Absolute Error and elevates the scientific evidence base of forensic medical expertise to a fundamentally new level of reliability.