SELECTING AN EFFECTIVE METHOD OF COLOR NORMALIZATION FOR HISTOLOGICAL IMAGES OF INTESTINAL TISSUES IN DEEP LEARNING MODEL DEVELOPMENT
D. V. Robota, Б. С. Бурлака · Актуальні проблеми сучасної медицини Вісник Української медичної стоматологічної академії · 2025
The advancement of modern computer technologies opens new opportunities for the automated analysis of whole-slide histological images. This is made possible by digital pathology approaches and artificial intelligence methods, particularly machine learning and deep learning. One of the key challenges in this process is the significant variability in the color of histological images. This variability arises from different staining techniques, the characteristics of laboratory equipment used for scanning specimens, and the individual properties of patient tissues. These factors can significantly affect the accuracy of automated decision-support systems, histological image classification, and tissue structure segmentation. The absence of a unified color normalization standard complicates the application of deep neural networks in histological image analysis tasks. This study examines the selection of a color normalization method for intestinal tissue images in the development of deep learning models aimed at multiclass segmentation of structural and functional tissue components. The conducted research characterizes approaches to standardizing the features of whole-slide histological images, which can be used to improve the accuracy of histopathological analysis of intestinal tissues. Various normalization methods were evaluated, and the most suitable approach for this type of task was identified. The experimental results demonstrate the high efficiency of using the Reinhard Modified method, which ensures high segmentation quality metrics: Mean IoU: 0.7086, Mean Dice: 0.8279, Precision: 0.8321, Recall: 0.8241, Accuracy: 0.8241, F1-Score: 0.8279, Specificity: 0.9112. The obtained results confirm the potential of color normalization in digital pathology and automated histological image analysis. The application of standardized normalization methods can significantly improve the accuracy of artificial intelligence systems, which is crucial for medical diagnostics and research activities. The advancement of modern computational technologies has opened new opportunities for the automated analysis of whole-slide histological images. This progress is driven by digital pathology approaches and artificial intelligence (AI) methods, particularly machine learning and deep learning. One of the key challenges in this field is the significant variability in the color of histological images. This variability arises due to differences in staining techniques, variations in laboratory equipment used for specimen scanning, and the unique properties of patient tissues. Such factors can substantially impact the accuracy of automated decision-support systems, histological image classification, and tissue structure segmentation. The absence of a unified color normalization standard further complicates the application of deep neural networks in histological image analysis. This study examines the selection of an optimal color normalization method for intestinal tissue images in the development of deep learning models designed for multiclass segmentation of structural and functional tissue components. The article characterizes various approaches to standardizing whole-slide histological image features, which can enhance the accuracy of histopathological analysis. Multiple normalization methods were evaluated, and the most suitable approach for this task was identified. Experimental results demonstrate the high efficiency of the Reinhard Modified method, which achieved superior segmentation quality metrics: Mean IoU: 0.7086, Mean Dice: 0.8279, Precision: 0.8321, Recall: 0.8241, Accuracy: 0.8241, F1-Score: 0.8279, Specificity: 0.9112. The findings confirm the potential of color normalization in digital pathology and automated histological image analysis. Implementing standardized normalization methods can significantly improve the accuracy of AI-driven diagnostic systems, which is crucial for medical diagnostics and research applications.