Abujalambo Mahmoud, A.I.M.Lazam, N.A.B.Md.Jawarneh, M.Hilles, S.M.S.Altrad, A.2025-05-312025-05-3120251992-86452-s2.0-105001963334https://hdl.handle.net/20.500.14517/7970Magnetic Resonance Imaging (MRI) is a widely used, non-invasive method for medical imaging, particularly effective in visualizing soft tissues and identifying abnormalities like spine hemangiomas. One of the main challenges remains the low segmentation accuracy of skeletal MRI images. Spine hemangioma segmentation involves algorithmically identifying and localizing these tumors within MRI scans, a process crucial for accurate diagnosis and treatment planning. Although several segmentation methods exist, this paper introduces a U-Net-based approach, implemented in PyTorch and optimized with the Adam optimizer. This setup refines model weight adjustments and harnesses the full capabilities of a fully connected convolutional neural network (CNN) for precise semantic segmentation, including pixel-wise classification through an encoder-decoder structure. This U-Net architecture is versatile and adaptable to various analytical tasks across diverse applications. The model was trained on a substantial dataset spanning the three primary anatomical planes used in medical imaging—Axial, Coronal, and Sagittal without additional data augmentation. It achieved real-time segmentation with a remarkable accuracy of 94.13% and demonstrated strong performance metrics, including a Dice coefficient of 0.634 and Precision of 0.711, underscoring its robustness and potential clinical utility. This work highlights U-Net’s effectiveness in spine hemangioma segmentation and explores its matching capabilities, indicating promising potential for advancements in automated MRI analysis. © Little Lion Scientific.eninfo:eu-repo/semantics/closedAccessAccuracyConvolutional Neural Network (Cnn)Dice CoefficientMri Spine Hemangioma SegmentationPrecisionSemantic SegmentationU-Net ModelArticleN/AQ4103624682478