Attention U-Net untuk Segmentasi Perdarahan Otak Multi-Kelas pada Dataset CT Kepala

Abstract

Brain hemorrhage is a life threatening condition that requires rapid and accurate assessment of hematoma location and volume on head CT images. Manual delineation by radiologists is time-consuming and prone to inter observer variability, so automated segmentation methods are needed. This study investigates an Attention U-Net model for automatic brain hemorrhage segmentation on CT images. The dataset, obtained from Roboflow in COCO segmentation format, was converted into pixel-wise binary masks and divided into training, validation, and test sets. Pre-processing included resizing, normalization, and data augmentation to improve generalization. The Attention U-Net was trained using the Adam optimizer with a learning rate of 1×10⁻⁴ for 20 epochs, employing a combined Binary Cross Entropy and Dice loss. Quantitative evaluation on the test set yielded a test loss of 0.1935 and a Dice coefficient of 0.6696. These results indicate that Attention U-Net is a promising approach for assisting brain hemorrhage assessment and can be further improved with larger and more diverse datasets