Brain Tumor and Haemangioma Discrimination Framework Using Unsupervised Feature Selection

Nowadays, in medical imaging, discrimination of brain tumors and haemangiomas is a major challenge due to tumor heterogeneity, overlapping vascular structures, and limitations in existing segmentation approaches. To address these issues, a Brain Tumor and Haemangioma discrimination framework is proposed that integrates unsupervised feature selection with a multi-stage pipeline. First, the Dense Feature Reuse and Volumetric U-Net (DFR-V-Net) is employed to segment the 3D MRI scans, enhancing feature representation while reducing redundancy in tumor boundaries. Next, a Weighted Voxel Upscaling method is introduced to correct partial volume effects and improve voxel-wise resolution, facilitating more accurate separation of tumor and normal tissues. The enhanced volumetric features are then processed using a Contiguous Factor Disentangled Variance Autoencoder (CFD - VAE), which isolates vascularization and blood flow-based characteristics that differentiate haemangiomas from malignant tumors. Finally, a Binary Gaussian-Initiated Self- Organizing Map (SOM) classifier performs unsupervised discrimination of tumor categories by modeling Gaussian cluster priors across intensity gradients and vascular orientations. The experimental validation on the proposed framework achieved 99.7% accuracy, 99.8% precision, 99.7% recall, 99.5% F1 score and 98.4% Dice Similarity Coefficient (DSC), which is significantly better than other conventional methods. The integration of voxel-level correction, disentangled vascular feature learning, and Gaussian-based unsupervised classification collectively enhances diagnostic reliability. This proposed framework provides a clinically practical solution for the precise and rapid differentiation of brain tumors and haemangiomas in MRI-based diagnosis.