ABSTRACT
Brain tumors are one of the most devastating anomalies in the human brain, and it often leads to brain cancer with a very short life expectancy. Accurate brain tumor diagnosis is a challenge for radiologists and domain experts. Noise corruption and inhomogeneities in images complicate the process even further and place constraints on feature analysis, making the task time-consuming and error-prone. Intersections of biomedical image processing and artificial intelligence have blazed the road for automated computer-aided diagnosis in recent years, with encouraging outcomes. In this chapter, we leverage the use of deep learning and transfer learning for robust automated tumor diagnosis. The diagnosis pipeline includes image denoising, tumor detection, and tumor grade identification phases in order to aid doctors in detailed medical diagnosis and early abnormality detection. The denoising phase enhances the image quality by removing aberrations and maintaining important details, thereby ensuring that the analysis tools are not sensitive to noise corruption in images. A denoising convolutional autoencoder focused on brain magnetic resonance (MR) image denoising is presented in this chapter. Next, five deep neural network architectures are fine-tuned for tumor detection and classification of different types of tumor (glioma, meningioma, and pituitary). After rigorous experimentation on standardized brain MRI datasets, the presented architectures manage to achieve resilient results, thus signifying reliability in automated brain tumor diagnosis.
