IMPROVED ORAL CANCER CLASSIFICATION USING HYBRIDIZATION OF DEEP FEATURE AND CORRELATION-BASED FEATURE WEIGHT

This study presents a hybrid deep learning-based feature classification framework for binary cancer detection using medical images. The proposed method extracts heterogeneous deep features from ResNet18 and EfficientNet-B0, which are then concatenated to form a unified feature vector. To quantify feature relevance, the Pearson correlation between each feature dimension and the ground-truth labels is computed. These correlation scores are then used to weight each feature dimension, producing a correlation-enhanced representation. The weighted feature matrix is then used to train multiple classifiers, including Random Forest, SVM, XGBoost, Extra Trees, and Logistic Regression.