Abstract
Background: Proper and interpretable brain tumor classification
is essential in making a successful clinical decision in neurooncology.
Automated approaches are potentially promising, but a lack of transparency
in decision-making is usually an obstacle to clinical implementation.
Objective: The proposed research developed and evaluated a convolutional
neural network (CNN) model in the context of automatic brain tumor classification
using magnetic resonance images (MRI) with a particular focus on
a high-performing model and visualizing predictions using Gradient-weighted
Class Activation Mapping (Grad-CAM). Methods: It utilized a dataset of
7,023 MRI scans as a sample, which was divided into glioma, meningioma, pituitary
tumors, and no-tumor. Preprocessing of the data was done by normalizing
and resizing, and stratifying into training, validation, and test subsets. The
suggested CNN has been compared with the state-of-the-art transfer-learning architectures,
such as VGG16, MobileNetV2, and DenseNet121. Results: The
proposed CNN had the highest predictive accuracy of 94.75%, precision of
94.99%, recall of 94.75%, and an F1-score of 94.82%, and better than all the
transfer-learning baselines. Moreover, Grad-CAM visualizations have always
identified tumor-specific areas in the images, confirming the clinical plausibility
of the model decisions. Conclusions: These results highlight the possibility
of high-performance CNN-based classification used in conjunction with
explainable AI to provide effective and high-quality diagnostic support that
is accurate, dependable, and explainable by clinicians. The future research
will explore the concept of multi-modal MRI integration, 3D architecture, and
privacy-preserving deployment schemes in the context of real-life healthcare
applications
is essential in making a successful clinical decision in neurooncology.
Automated approaches are potentially promising, but a lack of transparency
in decision-making is usually an obstacle to clinical implementation.
Objective: The proposed research developed and evaluated a convolutional
neural network (CNN) model in the context of automatic brain tumor classification
using magnetic resonance images (MRI) with a particular focus on
a high-performing model and visualizing predictions using Gradient-weighted
Class Activation Mapping (Grad-CAM). Methods: It utilized a dataset of
7,023 MRI scans as a sample, which was divided into glioma, meningioma, pituitary
tumors, and no-tumor. Preprocessing of the data was done by normalizing
and resizing, and stratifying into training, validation, and test subsets. The
suggested CNN has been compared with the state-of-the-art transfer-learning architectures,
such as VGG16, MobileNetV2, and DenseNet121. Results: The
proposed CNN had the highest predictive accuracy of 94.75%, precision of
94.99%, recall of 94.75%, and an F1-score of 94.82%, and better than all the
transfer-learning baselines. Moreover, Grad-CAM visualizations have always
identified tumor-specific areas in the images, confirming the clinical plausibility
of the model decisions. Conclusions: These results highlight the possibility
of high-performance CNN-based classification used in conjunction with
explainable AI to provide effective and high-quality diagnostic support that
is accurate, dependable, and explainable by clinicians. The future research
will explore the concept of multi-modal MRI integration, 3D architecture, and
privacy-preserving deployment schemes in the context of real-life healthcare
applications
Keywords
Convolutional Neural Networks; Magnetic Resonance Imaging; Brain Tumor Classification; Transfer Learning; Explainable Artificial Intelligence