DeepCAI-V3: Improved Brain Tumor Classification from Noisy Brain MR Images Using Convolutional Autoencoder and Inception-V3 Architecture

Brain tumors are abnormal cell growths within the brain tissues, necessitating their early detection towards effective treatment. To achieve this, high-quality brain images via medical imaging techniques, such as Magnetic Resonance Imaging (MRI), are essential for accurate brain health diagnosis. However, such MR images are often noise-prone, due to potential interference from various sources, including scanner artifacts, patient motion and intensity variations during capture, processing and storage steps. This noise can significantly hinder accurate classification of brain tumors. Consequently, most state-of-the-art methods for brain tumor detection struggle to perform well with noisy MR images as their error margins widen. Therefore, accurate brain tumor classification of noisy MR images remains an open problem. In this paper, DeepCAI-V3, a two-stage deep learning architecture, is proposed for improved brain tumor classification from Noisy MR images (NMRI). First, a convolutional Autoencoder is employed to pre-process and denoise NMRI. Subsequently, Inception- V3architecture is used for feature extraction and classification. DeepCAI - V3model was validated on the publicly available Brain MRI and Brain Tumor MRI datasets which are augmented with zero- mean Gaussian noise. Evaluation results indicate that DeepCAI- V3outperforms state-of-the-art methods having achieved accuracy of 97% and 99.5% on Brain MRI and Brain Tumor MRI datasets respectively, demonstrating its robustness against noise artifacts in brain MR images.