An Efficient Image-Fusion and CANFIS-Based Framework for Automated Glioma Brain Tumor Segmentation in MRI

Abstract

Glioma is the most aggressive class of primary brain tumor and its early, accurate delineation on magnetic resonance imaging (MRI) is critical for treatment planning and patient survival. Manual delineation is time-consuming and subject to inter-observer variability, which motivates automated computer-aided detection. This paper proposes a computationally efficient pipeline that combines multimodal image fusion with an adaptive neuro-fuzzy classifier for glioma identification and segmentation. Two source MRI slices are first fused using a combination of the Discrete Wavelet Transform (DWT) and the Stationary Wavelet Transform (SWT), guided by eigenvalue-weighted sub-band arithmetic, to enhance the visibility of abnormal tissue relative to either input alone. The fused image is then processed with a bank of twenty multi-orientation, multi-scale Gabor kernels, and the maximal response at each pixel is retained to form a single texture-enhanced image. Three complementary descriptor families — DWT statistical features, Gray-Level Co-occurrence Matrix (GLCM) features and Local Ternary Pattern (LTP) features — are extracted from the Gabor-filtered image and passed to a Co-Active Neuro-Fuzzy Inference System (CANFIS) classifier that labels each slice as glioma or non-glioma. Detected glioma slices are further processed with a dilation-erosion (morphological) operator to isolate the tumor boundary. On 178 slices from the BRATS 2015 benchmark (64 glioma, 114 non-glioma), the proposed system attains 97.8% sensitivity, 99.4% specificity and 98.9% accuracy for tumor-region segmentation, and a 98.3% overall detection rate when the three feature families are combined, compared with 65.8-76.5% for any single or paired feature family. The proposed morphological segmentation stage also outperforms conventional region-growing and watershed segmentation by 5.9-8.3 percentage points in accuracy. Comparative evaluation against three recent state-of-the-art methods shows consistent gains of 1.6-4.2 percentage points in accuracy and 6.7-9.4 percentage points in detection rate, indicating that the proposed fusion-plus-CANFIS pipeline is a competitive, low-complexity alternative to deep convolutional approaches for glioma screening in resource-constrained clinical settings.