Effective MRI based Brain Tumor Detection using Modified U-Net Model

Abstract

In recent times, image processing is an emerging application, which provides more detailed information about images, especially in the medical field. Image segmentation is an important task in medical image processing. Numerous imaging modalities are used in the medical field such as histopathology, X-ray, etc. In that, the Magnetic Resonance Imaging (MRI) provides precise segmentation results compared to other imaging modalities. The manual segmentation of brain tumor is a tedious and time consumption process, so an automatic detection model is proposed in this manuscript based on deep learning techniques. Firstly, the brain samples are acquired from the BraTS 2020 dataset, and then the quality of the brain image is modified by using color normalization technique. In addition, the data augmentation is performed by using scaling, rotation, and flipping techniques that decreases the overfitting risks and improves the flexibility and variability of the deep learning model. Finally, the brain tumor segmentation is performed using a modified U-net model that includes dice loss function and morphological gradient function for an effective segmentation in the MRI brain images. The modified U-net model attained Dice Similarity Coefficient (DSC) of 93%, 92%, and 95% for different brain tumor types such as Whole Tumor (WT), Enhanced Tumor (ET) and Tumor Core (TC), which are better related to the U-Net model.