A Comprehensive Review of Hybrid Deep Learning-Metaheuristic Framework for Accurate Feature Selection

Early diagnosis of cancer is difficult due to high-dimensional data, redundancies of features, and conventional machine learning (ML) and deep learning (DL) models’ limited interpretability.  This research paper gives an overview of metaheuristic based deep learning frameworks for feature selection. To study the recent literature on deep learning models like CNN, LSTM, Vision Transformer integrated with optimization algorithms such as Grey wolf optimizer, Particle swarm optimization, Genetic Algorithms, Bat Algorithm a PRISMA-based systematic approach has been adopted. Results show that hybrid models outperform other models, with accuracy exceeding 97% in datasets like BreakHis, DDSM, and TCGA. Specifically, Binary GWO + CNN achieved accuracy as high as 98.5%. Despite their growing popularity, limitations still persist such as lack of interpretability, dataset bias, and computational complexity.  Applications of these frameworks in medical imaging, genomic analysis and personalized healthcare contribute towards improved diagnosis accuracy and decision-making in clinical settings.