Abstract:
This work aims to enhance brain tumor diagnosis accuracy by examining crucial aspects
of cleaning and filtering MRI datasets, emphasizing the novel integration of Royal filtering
with the VGG19 architecture. It employs advancements in medical image processing and
deep learning to address challenges in using MRI for detecting brain malignancies. The
process begins by acquiring diverse brain MRI datasets. A systematic cleaning protocol is
applied, including conversions to grayscale, Gaussian blurring, thresholding,
morphological opening, and largest shape extraction. Royal filtering, in both 16-color and
royal versions, is a crucial step. The dataset is split 80/20 into training and testing sets.
Models undergo training and testing to evaluate various deep learning architectures:
VGG16, VGG19, InceptionV3, Xception, ResNet152V2, MobileNetV2, EfficientNetV2L,
EfficientNetV2M, ResNet50, and Royal VGG19. Royal VGG19 is best with 98.91%
accuracy. Ablation research on VGG19 provides insights into each component's
functionality. The proposed BTV19 model combines optimal preprocessing and Royal
filtering with VGG19. The research establishes a new framework for precise brain tumor
diagnosis and contributes by examining preparation and filtering techniques' impact on
deep learning efficacy. Findings were assessed using confusion matrices, ROC-AUC
curves, and k-fold cross-validation. Experiments show the proposed BTV-19 model
exhibits stability and reliability in improving brain tumor diagnosis accuracy. This work
significantly advances medical image quality and deep learning applications, ultimately
improving healthcare outcomes.
