Revolutionizing Medical Imaging Diagnostics: Unveiling the Potential of CNNs, ViTs, and CapsNets in Brain MRI Analysis

Abstract

This thesis addresses the challenge of enhancing diagnostic accuracy and efficiency in medical imaging, with a focus on brain MRIs—a choice dictated by the complexity of the subject matter and project time constraints. Currently, the diagnosis process for imaging like X-rays and MRIs relies heavily on radiologists, which can extend the turnaround time for results to three days or more. By leveraging deep learning tools (i.e., CNNs, CapsNets, Vision Transformers), this study aims to pioneer the capabilities for real-time analysis in medical diagnostics, potentially reducing result times to near immediacy. This foundational effort involves training each model separately to assess their individual contributions. Through meticulous evaluation, including training plots, cross-validation, and AUC scores, the research has achieved a notable 98% accuracy in tumor classification, underlining the significant potential of each technology. CNNs excel in feature detection, CapsNets in preserving spatial hierarchies, and Vision Transformers in understanding global contexts. These findings highlight the integrated models’ potential to markedly improve the precision and speed of medical imaging diagnostics, laying the groundwork for their application in clinical environments. Looking forward, the amalgamation of these technologies promises a robust model capable of overcoming fitting issues and adeptly classifying unseen data, establishing a solid foundation for future advancements in medical image analysis.