Year: 2026 | Month: March | Volume: 13 | Issue: 3 | Pages: 427-434
DOI: https://doi.org/10.52403/ijrr.20260349
Analysis of the Effect of Principal Component Analysis on Brain Tumor Classification Performance in MRI Images Based on Texture and Deep Features Using Subspace K-Nearest Neighbour
Fitriyani Mus Mulyadi, Pandji Triadyaksa, Qidir Maulana Binu Soesanto
Department of Physics, Faculty of Science and Mathematics (FSM), Diponegoro University, Semarang, Indonesia.
Corresponding Author: Pandji Triadyaksa
ABSTRACT
Advancements in deep learning and computer vision have improved the accuracy of MRI-based brain tumor classification. Yet, manual analysis still faces limitations such as subjectivity and long processing times. This study compares classical texture features (LBP, HOG, GLCM), feature combinations, and deep features from MobileNet for brain tumor classification using the Subspace K-Nearest Neighbour method. All MRI images were preprocessed, including grayscale conversion, resizing, CLAHE, and intensity normalization. Classical features were extracted with optimized parameters, while MobileNet was employed as a pre-trained feature extractor without full fine-tuning. Dimensionality reduction was performed using Principal Component Analysis (PCA) to improve computational efficiency. Experimental results show that HOG achieved the highest accuracy among classical features (0.967), while the LBP_HOG combination slightly improved performance (0.969). Deep features from MobileNet achieved the highest accuracy (0.974) with lower computational time compared to end-to-end training. PCA reduced the number of features by 42–46% for high-dimensional features without significant accuracy loss, whereas low-dimensional features were sensitive to reduction. The optimal configuration was identified as the MobileNet_Features + PCA + Subspace KNN pipeline, which maintained high accuracy (0.972) while minimizing computational complexity. These findings highlight the effectiveness of combining deep features, dimensionality reduction, and ensemble KNN in MRI-based brain tumor classification, offering an optimal balance between performance and computational efficiency.
Keywords: Brain tumor classification, MRI, Texture features, Deep features, Principal Component Analysis, Subspace K-Nearest Neighbor
[PDF Full Text]