3D MEDICAL IMAGE SEGMENTATION WITH DEEP LEARNING METHODS

OpenAlex İngilizce

With advancements in technology, three-dimensional (3D) medical imaging has become vital in modern medicine, contributing to more accurate diagnosis, treatment planning, and personalized medicine. However, segmenting abdominal organs remains a challenging task due to anatomical variations, limited labeled data, and image noise. This study investigates the impact of deep learning-based architectures and preprocessing techniques on 3D organ segmentation using the publicly available Multi-Atlas Labeling Beyond the Cranial Vault (BTCV) dataset. To achieve this, 3D U-Net, UNETR, and SwinUNETR models were employed, and the effects of various preprocessing techniques and loss functions, including Dice Loss, Focal Loss, and Cross-Entropy Loss, were systematically analyzed. The findings reveal that combining Dice Loss with Cross-Entropy Loss significantly enhances segmentation performance. Additionally, preprocessing techniques improved segmentation accuracy by 1.19%, further optimizing model performance. Among the evaluated models, 3D U-Net achieved the highest overall segmentation performance, with an average Dice score of 0.8397, outperforming SwinUNETR and UNETR. These findings underscore the importance of selecting appropriate preprocessing methods and loss functions in 3D medical image segmentation. The results contribute to more precise and efficient medical image analysis, with potential applications in clinical decision support systems. Future research should focus on optimizing hybrid architectures, integrating advanced augmentation strategies, and expanding evaluation across multiple datasets to improve the robustness and real-world applicability of automated segmentation methods.