A Linear Algebra and Spatial Distribution-Principal Component Analysis-Guided Dimensions Reduction for Brain Tumor Detection Using Optimized Residual Group Channel and Space Non-Convolutional Graph Neural Network

Abstract

Accurate and early diagnosis of brain tumors through Magnetic Resonance Imaging (MRI) plays a critical role in improving patient outcomes, guiding treatment planning, and enhancing survival rates. However, reducing high-dimensional MRI data often leads to the loss of essential spatial and structural features, adversely affecting diagnostic precision. Furthermore, many existing methods struggle with generalizability across diverse clinical datasets, raising concerns of overfitting in real-world applications. To address these limitations, this study introduces a novel Residual Group Channel and Space Non-Convolutional Graph Neural Network with Aphid-Ant Mutualism optimization (RGCS-NCGNN-AAM) for efficient and clinically reliable brain tumor detection. MRI scans sourced from the BraTS2018 and Figshare datasets are initially pre-processed using a Locally-Adaptive Bitonic Filter (LABF), which enhances image clarity by reducing noise while preserving tumor boundaries—crucial for clinical interpretation. Next, Spatial Distribution–Principal Component Analysis (SD-PCA) is employed to reduce dimensionality while retaining critical spatial tumor characteristics. For precise tumor localization, the Linear Algebra with Steerable Transformer (LA-ST) method ensures segmentation fidelity by leveraging matrix operations and geometric attention mechanisms. The core detection task is performed by the RGCS-NCGNN, which classifies MRI images into tumor and non-tumor categories while maintaining spatial context. To optimize performance, Aphid-Ant Mutualism (AAM)-based hyperparameter tuning balances exploration and exploitation through biologically inspired cooperative dynamics. Clinically, the proposed model demonstrates remarkable diagnostic potential, achieving 99.94% accuracy on BraTS2018 and 99.91% on Figshare, with high precision and F1-scores, and minimal false positives—critical in avoiding unnecessary interventions. By maintaining diagnostic reliability across heterogeneous datasets and enhancing interpretability of tumor features, this approach supports radiologists and neuro-oncologists in making informed, data-driven clinical decisions. The integration of biologically inspired intelligence and deep learning makes this framework a promising tool for real-world brain cancer diagnosis.