Semantic segmentation of tumor from 3D Structural MRI using U-Net Autoencoder

Abstract

Automated semantic segmentation of brain tumors from 3D MRI images plays a significant role in medical image processing. Early detection of these brain tumors is highly requisite for the treatment, screening, diagnosis and surgical pre-planning of the anomalies. The physicians normally follow the manual way of delineation to process the diagnosis of tumors which is time consuming, requires too much knowledge of anatomy and is too much expensive. To resolve these limitations, convolutional neural network (CNN) based autoencoder model is proposed which performs automated segmentation of brain tumors from 3D MRI brain images. Several algorithms such as image normalization, image augmentation, image binarization are used for data pre-processing. Furthermore, autoencoder based U-Net architecture is developed to extract the key features of the tumor and train the model. Later on, the model is applied to the new 3D MRI brain images to test the accuracy of it by segementing the tumor region. The proposed model enables enhancing the performance and accuracy of semantic segmentation of brain tumor as compare to the other existing models. Applying the proposed method, the accuracy is obtained upto 96.06% considering the 66 subjects. Finally, this approach is a well-structured model for segmenting the tumor region from MRI brain images which may assist the physicians for providing therapy and better treatment to the patient.