Design of Bio-Inspired Metaheuristics for Medical Image Classification

Abstract

Breast cancer diagnosis relies fundamentally on histopathological examination of tissue samples; however, manual microscopic evaluation by pathologists is subjective, time-consuming, and prone to inter-observer variability. While deep learning has demonstrated remarkable success in medical image analysis, conventional single-architecture models exhibit inherent limitations. Convolutional neural networks excel at local feature extraction but fail to capture global contextual relationships, whereas Vision Transformers effectively model long-range dependencies but may be inefficient in representing fine-grained local details that are critical for histopathological interpretation. Moreover, optimizing deep learning models through hyperparameter tuning and feature selection remains computationally expensive and often suboptimal when relying solely on gradient-based optimization techniques. This thesis proposes HNet, a novel hybrid deep learning architecture that strategically integrates three complementary neural paradigms for histopathological image classification. HNet combines EfficientNet for efficient local feature extraction from tissue morphology, an Advanced Vision Transformer (AVT) for capturing global contextual relationships and long-range tissue patterns, and Capsule Networks (CapsNet) for explicitly modeling spatial hierarchies and part-whole relationships inherent in tissue organization. To address the challenges of high-dimensional feature spaces and suboptimal parameter tuning, a Genetic Algorithm (GA)-based feature selection mechanism is incorporated as a critical preprocessing step between the concatenated EfficientNet-AVT feature representations and the CapsNet input. This metaheuristic-driven optimization enables automatic identification of the most discriminative features while reducing dimensionality and computational complexity. Comprehensive experimental validation is conducted on the BreakHis dataset, encompassing binary (benign vs. malignant) breast cancer classification tasks. Detailed ablation studies quantify the individual contributions of each architectural component as well as the impact of GA-based feature selection on overall performance. Comparative evaluation against recent state-of-the-art methods demonstrates that the proposed HNet architecture achieves superior classification accuracy, sensitivity, specificity, and F1-score, establishing a new benchmark for breast cancer histopathology classification. The integration of metaheuristic-driven optimization with hybrid deep learning significantly enhances model robustness, generalization to unseen data, and computational efficiency compared to non- optimized approaches. Beyond empirical performance gains, this work demonstrates how hybrid metaheuristic, deep learning frameworks can be effectively integrated into clinical workflows by addressing real-world deployment constraints, including inference latency, computational resource utilization, and model interpretability. The proposed methodology provides a generalizable template for applying metaheuristic-optimized 5 hybrid deep learning to a wide range of medical image classification tasks, offering substantial potential for advancing automated computer-aided diagnosis systems and digital pathology.