Hybrid deep neural network with PCA based features optimization for enhancing brain tumor classification.

Area of Science:

• Medical Imaging and Artificial Intelligence
• Computational Pathology
• Machine Learning in Oncology

Background:

• Brain tumors present diagnostic challenges due to unpredictable growth and complex medical needs.
• Accurate classification of tumor types is crucial for effective treatment planning.

Purpose of the Study:

• To develop and evaluate a hybrid Principal Component Analysis (PCA) DenseNet121 convolutional neural network for improved brain tumor classification.
• To enhance classification accuracy for four types: Glioma, Meningioma, Pituitary, and No Tumor.

Main Methods:

• A hybrid model combining deep features from DenseNet121 with traditional texture descriptors: Gray Level Co-occurrence Matrix (GLCM), Local Ternary Pattern (LTP), and Color Coherence Vector (CCV).
• Preprocessing of MRI data using CCV with 27 discrete intensity bins to capture spatial connectivity and subtle intensity relationships.
• Dimensionality reduction of the feature space using Principal Component Analysis (PCA) applied exclusively to training data to prevent bias.

Main Results:

• Achieved a classification accuracy of 95.89% for brain tumor types.
• Precision, recall, and F1-score consistently maintained above 94%.
• Demonstrated effective reduction of overfitting through coinciding training and validation accuracy/loss curves and successful use of dropout normalization.

Conclusions:

• Integrating multimodal texture descriptors with deep features provides a comprehensive tumor representation, minimizing misclassification.
• The proposed hybrid model ensures stable learning patterns and reliable diagnostic performance across diverse clinical datasets.
• The method effectively addresses the challenges of brain tumor classification, offering a promising tool for medical diagnostics.