Metaheuristic-Optimized Deep Learning for Lung Cancer Detection: A Systematic Review of Convolutional Neural Network Approaches

Abstract

Lung cancer remains the leading contributor to the global cancer burden, accounting for the highest mortality rates among both men and women, and underscoring the urgent clinical need for automated, accurate, and computationally efficient diagnostic systems. Convolutional Neural Networks (CNNs) have transformed medical image analysis for lung cancer detection, yet their performance remains sensitive to hyperparameter configuration. Metaheuristic optimization algorithms (MHAs) offer a principled automated alternative for CNN hyperparameter tuning, and their integration with chaotic maps further enhances global search capability. This systematic review synthesizes 82 peer-reviewed studies published between 2019 and 2026, following PRISMA 2020 guidelines, and organizes findings along three dimensions: imaging modality (computed tomography, chest X-ray, and histopathology), optimization strategy (swarm-based, evolutionary, chaotic-enhanced, transit search-based, and Particle Swarm Optimization [PSO]-hybrid), and performance benchmarks. The review documents mathematical formulations of 12 chaotic map variants, comparative accuracy benchmarks, and the progression from standard CNNs to sinusoidal chaotically enhanced transit search-based frameworks and, most recently, XAI-integrated lightweight designs. Peak accuracy of 98.88% for malignant classification was reported for the Sinusoidal Chaotic Transit Search Optimization Algorithm-based CNN (STSOA-CNN), while the 2025-2026 cohort extended the CT benchmark to 99.99% using Comprehensive Learning Particle Swarm Optimization. Meta-analytic aggregation across 23 quantitatively reported studies reveals a mean accuracy of 97.1% (SD = 2.3%) for MHA-optimized CNN approaches, compared with 93.6% (SD = 3.1%) for unoptimized baselines. Persistent limitations, including dataset heterogeneity, class imbalance, absence of external validation, and limited representation of resource-constrained settings, were identified, and a structured research agenda for the field is presented.