MBTC-Net: Multimodal brain tumor classification from CT and MRI scans using deep neural network with multi-head attention mechanism
S. Kar, Pawan Kumar Singh
Abstract
Brain tumors pose a singularly formidable threat in contemporary healthcare due to their diverse histological profiles and unpredictable clinical behavior. Their spectrum ranges from slow-growing benign tumors to highly aggressive malignancies in sensitive anatomical locations. This necessitates an intensified focus on their pathophysiology and demands precise characterization for patient-specific therapeutic solutions. Techniques to correctly identify brain tumors using artificial intelligence are often employed for addressing segmentation and detection tasks; however, the lack of generalizable results hinders medical practitioners from incorporating them into the diagnostic process. Predominantly reliant on Magnetic Resonance Imaging, research on other imaging methods like Positron Emission Tomography & Computed Tomography, is scarce due to a dearth of open-access datasets. Our study proposes a robust MBTC-Net framework by leveraging EfficientNetV2B0 for extracting high-dimensional feature maps, followed by reshaping into sequences and applying multi-head attention to capture contextual dependencies. After reintroducing the attention output into a spatial structure, we perform average pooling before transitioning to dense layers, enhanced with batch normalization and dropout. The model is fine-tuned with the Adamax optimizer to classify various kinds of brain tumors using softmax from T1-weighted, T1 Contrast-Enhanced, & T2-weighted MRI sequences and CT scans. To reduce the risk of overfitting, measures such as stratified 5-fold cross-validation have been extensively implemented across 3 open-access Kaggle datasets, obtaining 97.54% (15-class), 97.97% (6-class), and 99.34% (2-class) accuracies, respectively. We have also applied Grad-CAM to decipher and visually analyze the predictions made by this framework. This research underscores the need for multimodal training of CT scans and MRI sequences for deploying a sturdy framework in real-time environments and advancing the well-being of patients.