Hybrid deep neural network with PCA based features optimization for enhancing brain tumor classification.
Authors
Affiliations (4)
Affiliations (4)
- Department of Information Technology, College of Technology, Govind Ballabh Pant University of Agriculture and Technology Pantnagar, Pantnagar, Uttarakhand, India. [email protected].
- Department of Technical Education Uttar Pradesh, Kanpur, India.
- Medical Physics and Informatics Laboratory of Electronic Engineering, National Kaohsiung University of Science and Technology, Kaohsiung, Taiwan.
- Department of Statistics, College of Natural and Computational Science, Mizan-Tepi University, Tepi, Ethiopia. [email protected].
Abstract
Brain tumors have been an important medical concern. This is primarily due to their growth patterns have been hard to predict and their medical needs have been complicated. This work proposes emphasis on a hybrid PCA DenseNet121 convolutional neural network mechanism. The model is intended to improve classification accuracy among four specific classifications: Glioma, Meningioma, Pituitary, and No Tumor. The model incorporates deep features with traditional texture descriptors to obtain an entire set of features. Specifically, it combines the Gray Level Co-occurrence Matrix (GLCM) and Local Ternary Pattern (LTP) with the Color Coherence Vector (CCV). Experimental integrity was a primary focus of this study. To preclude data leakage, the feature extraction and optimization pipeline was strictly partitioned between the training and testing datasets. The implementation of the CCV is methodologically supported by a specific preprocessing stage. In this stage, MRI intensity distributions are quantized into 27 discrete bins. This process classifies pixels as coherent or incoherent based on their spatial connectivity. This method captures subtle relationships between intensity clusters. These associations significantly improve the structural data provided by GLCM as well as LTP. Principal Component Analysis (PCA) has been employed to deal with the resultant feature space in many dimensions. This method preprocesses the feature vector beforehand to the training phase. The PCA transformation has been almost exclusively adapted to the training data, giving them highly significant. This approach eradicates anticipating bias despite ensuring that the model has applicability across diverse situations. The proposed approach has produced an accuracy of 95.89% with regard to classification. This performance has been confirmed by the measurements for the precision, recall, and a F1 score that have retained the same upon 94%. The findings indicate that incorporating multimodal descriptors alongside deep features provides a comprehensive representation of tumor features. This integration minimizes misclassification. It also ensures stable learning patterns throughout the diagnostic process. Additionally, the training and validation processes for each accuracy and loss constantly coincide. This indicates that the model performed an effective task of reducing down on too much excessive overfitting. The elevated degree of performance has been caused by the successful use of dropout normalization. After that, systematic information shuffling enhances the learning process more reliable with an extensive variety of clinical data sets.