USE OF U-NET WITH DIFFERENT DEPTHS FOR MENINGIOMA, GLIOMA AND PITUITARY TUMOR SEGMENTATION

• Main Author: Adiyono Cahyo, Nugroho
• Format: Final Project
• Language: Indonesia
• Online Access: https://digilib.itb.ac.id/gdl/view/79998
• Institution: Institut Teknologi Bandung

Brain tumor is an abnormal cell in brain organ that can develop over time. Segmentation can be used to help monitor tumor development also its location. Deep learning can be used to do segmentation of brain tumor. U-Net is one of the deep learning architecture that can be used for segmentating brain tumor. U-Net’s depth can be modified. U-Net’s depth can affect segmentation performance. Deeper U-Net performs higher F1-score but no significant improvement in the last three levels. Improvement occurs from level 1 to level 2 about 17.9%, level 2 to level 3 about 3.6%, level 3 to level 4 about 0.7%, and level 4 to level 5 about 1.1%. For meningioma, intrinsic factors affecting segmentation performance include circularity and mean intensity. In pituitary tumors, the features affecting segmentation performance are standard deviation of intensity. In glioma, the features affecting segmentation performance include circularity, standard deviation of intensity, and maximum intensity. U-Net’s failure to segmentate images that has inconsistency of ground truth, tumor intensity characteristics, quality of image acquisition. U-Net also can not learn the symmetry of brain to help segmentation and distinguish pituitary tumor and its surrounding.