Automated Brain Tumor Detection and Classification System

The accurate rapid and non-invasive detection of central nervous system malignancies remains one of the most critical challenges in contemporary neurology. Magnetic Resonance Imaging serves as the undisputed gold standard for visualizing these complex pathology however, the manual interpretation of multi parametric volumes is cognitively exhaustive and severely bottle necked by a global shortage of specialized anesthesiologists. While recent paradigms in deep learning have achieved unprecedented diagnostic accuracy they of ten do so at the cost of prohibitive computational complexity. Foundational work established the current state of the art on a comprehensive four class brain classification task using a fine tuned object detection architecture. However this apex performance is inextricably linked to extreme hardware requirements specifically reliance on dedicated graphics processing units with massive memory reserves. Such hardware dependency renders the deployment of this model virtually impossible in the vast majority of primary healthcare facilities particularly within low and middle income countries. To bridge this critical chasm this thesis presents a highly optimized universally central processing unit deploy able alternative. Our proposed system is engineered upon a highly efficient lightweight backbone strategically retaining advanced spatial attention mechanisms contextual pooling and bidirectional feature networks. The proposed system achieves a highly competitive ninety four percent overall accuracy while enabling under two second inference per image. Furthermore we extend the base architecture with a complete clinical workflow including a robust relational patient database a novel quantitative longitudinal tumor growth tracking engine and an automated diagnostic re- port generator all integrated into a clinical web dashboard.