JIITA, Vol.10 no.3 pp.1388-1405 (2026), DOI: 10.22664/ISITA.2026.10.3.1388
Sang Suh and Ozlem Deveci
Abstract. Brain tumor segmentation in MRI scans is crucial for accurate diagnosis, treatment planning, and monitoring tumor progression. Manual segmentation is timeconsuming and prone to errors, while traditional deep learning models often struggle with small or low-contrast tumor regions and precise boundary delineation [27]. In this research, we propose an enhanced brain tumor segmentation approach using a hybrid TransUNet architecture combined with a boundary-aware loss function adopted from [28]. The model leverages the convolutional encoder of U-Net for local features and a Transformer module
for capturing global context. The boundary-aware loss, which combines Dice, CrossEntropy, and edge-focused terms, improves segmentation accuracy, especially for the clinically important enhancing tumor regions.
Keywords; brain tumor segmentation, enhancing tumor segmentation, boundarydelineation, TransUNet architecture
Fullpaper: