A Comparative Analysis on Predicting Brain Tumor from MRI FLAIR Images Using Deep Learning

Abstract

It is still challenging to differentiate between normal cells and tumor demarcation in everyday clinical practice. With the use of the FLAIR modality known as Fluid Attenuated Inversion Recovery, a medical professional can learn more about tumor infiltration. Because the preponderance of the cerebrospinal fluid effect can be suppressed by the FLAIR modality. Moreover, one of the advantages of using FLAIR images is that they can be used for both 3D and 2D medical imagery. Therefore, this paper explores the idea of assessing and predicting brain tumors by implementing several types of deep learning CNN architectures, such as VGG16, ResNet50, DenseNet121 and others in a user-friendly functional U-Net architecture. The flexibility of using different pre-trained neural network models in a single architecture is the key advantage of our U-Net architecture. Hyperparameters of the architecture are adjusted and fine-tuned for the segmentation process in order to extract the core features of the tumor contour according to our problem. Having said that, this study's segmentation result on the dice similarity coefficient is 0.9165, 0.9175, 0.9137 and 0.9148 in the BraTS 2018, 2019, 2020 and 2021 datasets respectively.