INVESTIGATING THE IMPACT OF ADAM AND SGDM OPTIMIZERS ON U-NET TRAINING FOR SEGMENTATION OF EPITHELIAL COMPARTMENTS IN NORMAL INTESTINAL TISSUE

The integration of computer technologies into medical practice is fundamentally transforming approaches to histopathological data analysis, positioning digital pathology as an essential component of the diagnostic workflow. The use of Whole Slide Images (WSIs) in combination with deep learning methods enables the automation of complex tasks, including tissue structure segmentation. A key challenge in this field is achieving high accuracy in the recognition of small morphological elements within high-resolution images, which critically depends on neural network architecture and training parameters. These factors directly influence the reliability of automated decision-support systems, particularly in the analysis of intestinal tissue, where precise differentiation between epithelial and stromal components is essential. The absence of a well-founded strategy for selecting optimization algorithms may result in the loss of diagnostically relevant information and the generation of segmentation artifacts. This study investigates the impact of different optimization algorithms within the U-Net architecture on the accuracy of pixel-level classification of epithelial subcomponents (crypts and surface epithelium) in normal intestinal tissue. In the present study, the performance of the Adam and stochastic gradient descent with momentum (SGDM) optimizers was evaluated for training a U-Net convolutional neural network with a ResNet50 encoder. A comparative analysis of the resulting models was conducted using a histological image dataset to determine the optimal configuration for segmenting this type of morphological structure. The experimental findings demonstrated a clear advantage of the Adam optimizer, which ensured superior segmentation performance and greater training stability. Specifically, the Adam-based model achieved a Global Accuracy of 0.96 and a Specificity of 0.93, representing a 14.81% relative increase in Specificity compared to the SGDM optimizer. These results highlight the potential of adaptive optimization methods in digital pathology tasks requiring high-resolution segmentation. The use of the Adam algorithm reduced the number of false-positive predictions in stromal regions, thereby improving the reliability of morphometric analysis. This enhancement is particularly relevant for both fundamental and clinical oncology research.