Deep learning based brain tumor segmentation

Brain tumor segmentation aims to separate healthy tissue from tumor sites. Today, brain tumor segmentation is performed manually by radiologists. In recent years, automated computer-aided systems implemented using deep learning methods have gained great momentum by replacing manual detection systems. More than half of brain tumor types are gliomas. Manually segmenting glioma tumors using brain MRI images is quite challenging. The reason for this is; glioma tumors are different in shape, size, volume, location and appearance, and it is difficult to distinguish between the two because healthy tissues and tumorous tissues are similar in structure. As a data set, brain MRI images of 2019 in the BRATS competition held every year were used. In this study, unlike other studies, appropriate modalities were used to see different parts of the tumor more clearly and MR images were cropped according to tumor sizes. 3 model trainings were carried out separately for complete tumor segmentation, expanding + necrosis and expanding tumor segmentation. As a result of the trainings conducted, compared to the studies in the literature, this study was among the pioneering studies as the 2nd in complete tumor segmentation, 2nd in necrosis + expanding tumor segmentation, and 1st in expanding tumor segmentation.