ET-WOFS Metaheuristic Feature Selection Based Approach for Endometrial Cancer Classification and Detection

Abstract

Endometrial Cancer (EC), also referred to as endometrial carcinoma, stands as the most common category of carcinoma of the uterus in females, ranking as the sixth most common cancer worldwide among women. This study introduces a Machine Learning-Based Efficient Computer-Aided Diagnosis (ML-CAD) state-of-the-art model aimed at assisting healthcare professionals in investigating, estimating, and accurately classifying endometrial cancer through the meticulous analysis of H&E-stained histopathological images. In the initial phase of image processing, meticulous steps are taken to eliminate noise from histopathological images. Subsequently, the application of the Vahadane stain normalization technique ensures stain normalization across histopathological images. The segmentation of stain-normalized histopathological images is executed with precision using the k-NN clustering approach, thereby enhancing the classification capabilities of the proposed ML-CAD model. Shallow features and deep features are extracted for analysis. The integration of shallow and deep features is achieved through a middle-level fusion strategy, and the SMOTE-Edited Nearest Neighbor (SMOTE-ENN) pre-processing technique is applied to address the sample imbalance issue. The identification of optimal features from a heterogeneous feature dataset is conducted meticulously using the novel Extra Tree-Whale Optimization Feature Selector (ET-WOFS). For the subsequent classification of endometrial cancer, a repertoire of classifiers, including k-NN, Random Forest, and Support Vector Machine (SVM), is harnessed. The classifier that incorporates ET-WOFS features demonstrates exceptional classification outcomes. Compared with existing models, the outcomes demonstrate that a k-NN classifier utilizing ET-WOFS features showcases remarkable outcomes with a classification accuracy of 95.78%, precision of 96.77%, an impressively low false positive rate (FPR) of 1.40%, and also a minimal false negative rate (FNR) of 4.21%. Further validation of the model's prediction and classification performance is evaluated in terms of the AUC-ROC value and other metrices. These presented assessments affirm the model's efficacy in providing accurate and reliable diagnostic support for endometrial cancer.