Brain Tumor Detection and Segmentation Using Deep Learning

Abstract

Brain tumor detection and segmentation refer to the processes of identifying the presence of brain tumors in medical imaging data and accurately delineating the tumor boundaries within the images. These processes involve classifying the pixels within the image as tumor or non-tumor regions. Accurate segmentation is essential for quantifying tumor size, analyzing tumor characteristics, monitoring tumor progression, and assisting in treatment planning. Brain tumor detection and segmentation are widely used techniques in brain magnetic resonance imaging (MRI) analysis for quantifying and visually representing the anatomical structures of the brain, detecting pathological regions, planning medical procedures, and guiding treatments. Due to the wide variety in the size, shape, location, and appearance of brain tumors, brain tumor detection and segmentation are challenging tasks that needs extensive experimental investigation. Deep Neural Network algorithms like UNet outperform many semantic detection and segmentation tasks. However, Unet’s shallow networks limit the model’s ability to extract abstract and global contextual features. Therefore, this thesis presented two improved Unet architectures, that are DeepUnet as a baseline and Residual Unet as a proposed model, which capture multiscale features to enable the models to extract both local and global contextual features. The proposed method is hypothesized to capture important feature of non-uniform shapes, sizes and location from MRI images. Residual Unet, enhances the Unet architecture by incorporating a residual and skip connection module to handle both local and global features, as well as increasing the depth of the encoder network to extract more abstract features. These improvements resulted in higher detection and segmentation accuracy, providing clinicians with more accurate diagnostic and treatment planning information for brain tumor patients. Experimental results showed that the proposed model Residual Unet has achieved an accuracy of 99.33% on Brats 2020, and 98.75% on Brats 2019.On the other hand Deep Unet which is the baseline for the proposed model has achieved an accuracy of 99.11% on Brats 2020 and 98.58% on Brats 2019.So Residual model has outperformed better than baseline model which is Deep Unet in terms of performance. These findings indicate that the proposed models can greatly improve and facilitate brain tumor detection and segmentation.