Improved Classification of Different Brain Tumors in MRI Scans Using Patterned-GridMask

Abstract

Continuous advancements in deep learning are affecting various research areas, especially research on applications in the medical sector. A computer-aided diagnosis system that utilizes deep learning is used for classifying and detecting brain tumors in magnetic resonance imaging. Regarding brain tumors, the main diagnostic indicators are patient symptoms and outcomes of magnetic resonance imaging. Frequent changes in the symptoms of these tumors have raised serious concerns about potential misdiagnoses. Implementing computer-aided diagnosis systems can support diagnostic methods that rely on the visual assessments of physicians, potentially reducing misdiagnosis rates. In this study, we propose an enhanced computer-aided diagnosis algorithm that is optimized for brain tumor classification. We removed noise from the magnetic resonance imaging results by applying Gaussian filters, and we employed GridMask to improve the generalization performances of the deep learning models. Then, we applied Patterned-GridMask, which is a method we proposed to reduce the issue of brain tumors being obscured by standard GridMask. Under the application of Patterned-GridMask, a performance improvement of up to 6% was demonstrated across the four deep learning models used in the experiments: ViT-B/16, MaxViT-B, TresNet-M, and EfficientNetV2-M, with the highest performance being represented by an accuracy and F1-score of 97.74% and 97.75%, respectively. Using the proposed computer-aided diagnosis system, improved diagnosis results can be obtained, thereby resulting in more accurate rates of early detection, better patient outcomes, and more appropriate treatment selection.