Бучко, ОленаПлахотна, Дар’я2026-02-022026-02-022025Бучко О. А. Порівняння архітектур нейронних мереж для сегментації пухлин мозку / Бучко О. А., Плахотна Д. О. // Наукові записки НаУКМА. Комп'ютерні науки. - 2025. - Т. 8. - С. 108-112. - https://doi.org/10.18523/2617-3808.2025.8.108-1122617-38082617-7323https://ekmair.ukma.edu.ua/handle/123456789/38237https://doi.org/10.18523/2617-3808.2025.8.108-112Image segmentation plays a crucial role in medical diagnostics, where precise identification of tumor boundaries is essential for treatment planning and prognosis. However, accurate segmentation remains a challenge due to the complexity of anatomical structures and the limited availability of annotated data. Traditional methods are not robust to variability in medical images, which often results in inconsistent and inaccurate outcomes. This paper investigates the effectiveness of modern neural network-based architectures for brain tumorsegmentation using MRI data. The primary goal of the study is to evaluate and compare the accuracy of several convolutional segmentation models under identical training conditions, and to examine whether self-supervised pretraining can improve segmentation quality in cases with limited labeled samples. The research is based on the BraTS2020 dataset, which contains multi-modal MRI scans with manual annotations of glioma subregions. Five segmentation models were trained: FCN, FPN, PSPNet, DeepLabv3+, and Attention U-Net. All models were evaluated using the Dice Similarity Coefficient. The best result was achieved by Attention U-Net with a mean Dice score of 0.842. A self-supervised learning (SSL) strategy was further applied to pretrain the encoder of this model using a rotation prediction task, which increased the Dice score to 0.869. The findings confirm that neural network-based methods provide higher segmentation quality compared to classical approaches. More importantly, the integration of SSL enables performance improvements without requiring additional labeled data. This is particularly valuable in the medical field, where collecting expert annotations is expensive and time-consuming. This article demonstrates that high-quality segmentation of brain tumors is possible even under limited supervision, provided that suitable architectures and training strategies are selected. The presented approach can be adapted for other medical image segmentation tasks and may support the development of practical clinical decision support systems.У роботі досліджено ефективність сучасних архітектур глибокого навчання для сегментації медичних зображень у задачі виявлення пухлин мозку. Проведено порівняльний аналіз моделей FCN, DeepLabv3+, PSPNet та Attention U-Net. Окрему увагу приділено впливу попереднього самоконтрольованого навчання на якість сегментації в умовах обмеженої кількості розмічених даних. Для оцінювання результатів використано метрику подібності.ukсегментація зображеньглибоке навчаннянейронні мережіAttention U-NetBraTS2020самоконтрольоване навчаннястаттяimage segmentationneural networksAttention U-Nself-supervised learninBraTS2020Article