Multiscale-Based Multi-Modal Image Classification of Brain Tumor Using Deep Learning Method

Magnetic Resonance Imaging (MRI) serves as a widely employed diagnostic method for evaluating glioma-type brain
lesions. This study introduces an automated segmentation approach for glioma-type tumors in MRI scans, employing UNET Convolutional Neural Networks (CNN) with compact 3x3 kernel sizes. The adoption of small kernels serves to
combat overfitting in deep neural networks and reduces the computational complexity associated with numerous
network weights. To address the significant spatial and structural variations within brain tumors, a deep learning
semantic segmentation technique known as the Multi-Scale Multimodal Convolutional Neural Network (SSMCNN) is
employed, accommodating multiple MRI modalities. The primary objective of this methodology is the precise
identification and separation of tumor categories by evaluating each individual pixel within the images. Furthermore, employing a classification approach based on patches enhances the effectiveness of semantic segmentation. The method utilizes a deep convolutional network, the Multiscale U-NET, to categorize multimodal images into three distinct scale segments through detailed analysis at the pixel level. By amalgamating these methodologies, the proposed approach aims to achieve precise and consistent image segmentation, a crucial aspect of effective clinical evaluations in glioma diagnosis.