Analysis of stomach cancer with artificial intelligence techniques

In this thesis, the performance of different feature extraction, dimensionality reduction, and classification methods was investigated for the automatic classification of histopathological images of gastric tumors. The study utilized histopathological images belonging to three classes: normal, benign, and malignant. For each class, 30 images were used for training and 30 images for testing, ensuring class balance and enabling a fair performance comparison. Feature extraction was performed using Accelerated Robust Features (ARF) and Maximally Stable Extremal Regions (MSER). The extracted features were transformed using Principal Component Analysis (PCA) and Generalized Discriminant Analysis (GDA) approaches in order to obtain a more compact and discriminative representation. The reduced feature sets were evaluated using Decision Trees, Random Forest, Artificial Neural Networks, and Naive Bayes classifiers. The number of features was gradually increased from 1 to 50 to analyze the behavior of the methods under different feature dimensionalities. The experimental results demonstrated that combining class-aware dimensionality reduction techniques with powerful classifiers significantly improves classification accuracy. In particular, MSER–GDA-based approaches provided more effective discrimination between benign and malignant tumor classes. The findings indicate that the proposed methodology has strong potential to serve as a reliable component in computer-aided diagnosis systems.