Integrating image processing with deep convolutional neural networks for gene selection and cancer classification using microarray data

Microarray technology has revolutionized cancer genomics by enabling the simultaneous analysis of thousands of gene expressions, providing critical insights into gene regulation and disease mechanisms. However, the inherent challenges of high-dimensionality, noise, and sparsity in microarray data demand robust analytical approaches. Image processing techniques further enhance this analysis by extracting meaningful patterns from histological and microarray-derived visual data, aiding in biomarker discovery and classification. This study presents a novel framework leveraging deep neural networks for gene selection and cancer classification using microarray data, addressing the challenges of high dimensionality, noise, and sparsity. The proposed Gene-Optimized Neural Framework (GONF) integrates the Minimum Redundancy Maximum Relevance (mRMR) gene selection method with a deep Convolutional Neural Network (CNN) for effective feature selection and classification. By optimizing hyperparameters and employing advanced preprocessing techniques, the framework enhances computational efficiency and accuracy. Experiments were conducted on TCGA and AHBA datasets, utilizing metrics such as accuracy, precision and recall for evaluation. The GONF outperformed other methods, achieving a classification accuracy of 97% on the TCGA dataset and 95% on the AHBA dataset. The framework demonstrated significant reductions in false positive and false negative rates, improving cancer subtype predictions and providing biologically interpretable results. The findings highlight GONF's robustness and adaptability, paving the way for its application in other genomic studies and clinical settings.