Brain Tumor Segmentation and Classification Using Neural Networks Based on Selected Features

Abstract

The human brain, a highly intricate organ, governs the entirety of physiological functions. The emergence of brain tumors, characterized by the anomalous and unregulated proliferation of brain cells, both within and outside the cranial cavity, presents a multifaceted challenge. These tumors manifest diversely in terms of their spatial distribution, morphology, and radiological attributes. Brain tumor segmentation entails the precise demarcation of pathological tumor tissue from normal brain constituents, while classification pertains to the discernment of the specific tumor subtype based on its distinctive features. The accuracy of brain tumor segmentation holds paramount significance in the realms of diagnosis, patient monitoring, and treatment strategizing, particularly for individuals afflicted by cerebral malignancies. This computational challenge is extremely complex since it lies at the intersection of computer vision and medicine. The most common method for evaluating brain malignancies is magnetic resonance imaging (MRI). Nonetheless, the manual segmentation and classification of 3D MRI images impose an arduous and time-intensive burden, contingent upon operator proficiency, leading to variable outcomes. In light of these challenges, the imperative arises for the development of a dependable, fully automated method for brain tumor segmentation and classification, offering efficiency and consistency in delineating tumor subregions. Convolutional Neural Networks (CNNs), a representative of deep learning techniques, have surpassed earlier machine learning paradigms in this endeavor by successfully completing the challenging process of segmenting brain tumors. Within this context, we introduce a deep learning-based framework tailored for the segmentation of brain tumors from multi-modal MRI scans. This innovative framework draws inspiration from two prominent architectural paradigms, namely the U-Net and residual network, further enhanced with attention mechanisms. The embedded attention gates facilitate automatic focalization on structures of varying dimensions and shapes while suppressing irrelevant regions. Our method underwent rigorous evaluation on the BRATS 2015 dataset, comprising 220 High-Grade Glioma (HGG) cases and 54 Low-Grade Glioma (LGG) cases. The results exhibited Dice scores of 0.53, 0.73, and 0.61 for enhancing tumor, whole tumor, and tumor core segmentation, respectively, on the BRATS 2015 test data, affirming the efficacy of our approach.