Investigation of Performance of Different U-Net Architecture for Brain Tumor Segmentation on Lower-Grade Glioma (LGG) Dataset

A brain tumor is a type of cancer, or more precisely, an abnormal mass that arises from uncontrolled cell division and can be benign or malignant. Early detection of a tumor increases the chance of survival. In recent years, magnetic resonance imaging (MRI) of the brain has been widely used in neurooncology clinics for imaging the anatomy and structure of the brain, diagnosis, treatment planning, and tumor monitoring after treatment due to its high contrast for soft tissues compared to other medical imaging methods. In addition to having high contrast for soft tissues, this method has none of the disadvantages of radioactive radiation for humans. Therefore, accurate segmentation of brain tumor lesions in MRI images is one of the important challenges in medical image processing. In this paper, three methods - U-Net, Nested U-Net and Inception/U-Net - are employed. Finally, the performance of the models for the segmentation process is compared based on standard evaluation criteria such as precision, recall, F۱ score and IoU, and it is found that the U-Net mode has the best performance, achieving the highest value in all the criteria.

کلیدواژه ها:

Cancer ، Brain Tumor Segmentation ، U-Net Architecture ، Magnetic Resonance Images ، Deep Learning