CANCER DIAGNOSIS FROM GENE EXPRESSION DATA USING FUZZY CLASSIFIERS AND DEEP LEARNING

Abstract

Microarray gene expression data poses a significant challenge in classification due to its small sample size and high dimensionality. In this thesis, we propose novel approaches for the classification of lung cancer subtypes using advanced techniques and algorithms. The first approach introduces the Fuzzy Min-Max (FMM) classifier, a neuro-fuzzy neural network rarely used for high-dimensional datasets. To enhance the accuracy and speed of FMM, we incorporate the Least Absolute Shrinkage and Selection Operator (LASSO) for optimal gene subset selection. Comparative analysis with other classifiers, including SVM, Random Forest, KNN, Naïve Bayes, and Logistic Regression, validates the superior performance of FMM LASSO in lung cancer classification. The second approach addresses the challenges of small sample sizes, high dimensionality, and class imbalance in cancer subtyping. Our proposed SMOTE-LASSO-DeepNet framework employs SMOTE for data balancing and LASSO for informative gene selection. The pruned and balanced training set is then fed into a DeepNet model with multiple hidden layers. Extensive testing on four different cancer gene expression datasets demonstrates the consistent superiority of our framework over existing methods. In the third approach, we tackle lung cancer diagnosis using gene expression data. Leveraging the Fuzzy Min-Max (FMM) classifier, specifically the general Fuzzy min-max (GFMM) and enhanced Fuzzy min-max (EFMM) models, we exploit fuzzy class definitions and hyperbox manipulation. LASSO is utilized for informative gene selection, and the performance is evaluated through hyperbox visualization and comparison with state-of-the-art methods. Empirical results showcase the exceptional performance of GFMM with LASSO, achieving validation accuracy of 98.04% and cross-validation accuracy of 94.06%. Collectively, these approaches contribute to the field of cancer diagnosis from gene expression data, offering novel solutions for small sample sizes, high dimensionality, and class imbalance issues. The proposed methodologies demonstrate superior performance compared to existing methods and highlight the potential of neuro-fuzzy systems, deep learning frameworks, and feature selection techniques in improving cancer classification accuracy.