Brain tumor prediction on MR images with semantic segmentation by using deep learning network and 3D imaging of tumor region

• Published in: Biomedical signal processing and control Vol. 66; p. 102458
• Main Authors: Karayegen, Gökay, Aksahin, Mehmet Feyzi
• Format: Journal Article
• Language: English
• Published: Elsevier Ltd 01.04.2021
• Subjects: 3D imaging; Deep learning network; Semantic segmentation; Semantic segmentation; 3D imaging; Deep learning network
• ISSN: 1746-8094; 1746-8108
• DOI: 10.1016/j.bspc.2021.102458

•An automatic tumor prediction method based on sematic segmentation.•257 BRATS image data set were used for training process on deep learning network.•The method involves semantic segmentation metrics with high prediction and accuracy ratios.•The method also comprises 3D imaging for detecting the exact tumor region. When it comes to medical image segmentation on brain MR images, using deep learning techniques has a significant impact to predict tumor existence. Manual segmentation of a brain tumor is a time-consuming task and depends on knowledge and experience of physicians. In this paper, we present a semantic segmentation method by utilizing convolutional neural network to automatically segment brain tumor on 3D Brain Tumor Segmentation (BraTS) image data sets that comprise four different imaging modalities (T1, T1C, T2 and Flair). In addition, our study includes 3D imaging of whole brain and comparison between ground truth and predicted labels in 3D. In order to obtain exact tumor region and dimensions such as height, width and depth, this method was successfully applied and images were displayed different planes including sagittal, coronal and axial. Evaluation results of semantic segmentation which was executed by a deep learning network are significantly promising in terms of tumor prediction. Mean prediction ratio was determined as 91.718. Mean IoU (Intersection over Union) and Mean BF score were calculated as 86.946 and 92.938, respectively. Finally, dice scores of the test images were showed significant similarity between ground truth and predicted labels. As a result, both semantic segmentation metrics and 3D imaging can be interpreted as meaningful for diagnosing brain tumor accurately.