This study focuses on improving the accuracy of lung cancer subtype classification by integrating machine learning feature selection with a deep learning model, namely, a Multi-Layer Perceptron (MLP). Using gene expression data from the TCGA database, we employed threshold variance and chi-square test feature selection, normalized the selected gene features using z-score normalization, and finally applied MLP for deep learning classification. Through comparisons with traditional machine learning models such as KNN, SVM, and Random Forest, we comprehensively examined multiple evaluation metrics, including accuracy, recall, precision, and F1 score. The results indicate that our MLP approach exhibits significant advantages over traditional methods in all aspects, particularly achieving a notable improvement in accuracy ranging from 2.6% to 12.8%. Overall, our deep learning classification method demonstrates outstanding performance in lung cancer subtype tasks, providing robust support for the application of deep learning in the biomedical field.