Abstract
OBJECTIVE: Breast cancer (BC) remains a significant global health issue, with an estimated 2.3 million new cases and approximately 670,000 deaths recorded in 2022 alone. Accurate and timely diagnosis from histopathological images is critical yet challenging due to the manual and labour-intensive nature of conventional pathology assessments and high inter-observer variability. To address these challenges, this study introduces LExNet, a bio-inspired lightweight ensemble model that integrates hybrid autoencoder-based feature extraction and Particle Swarm Optimization (PSO) for hyperparameter tuning. METHODS: The proposed model utilizes a robust ensemble of lightweight convolutional neural network (CNN) architectures, specifically MobileNetV3, EfficientNet-B0, ShuffleNet-V2, and SqueezeNet, to ensure computational efficiency and high accuracy. RESULTS: Experimental evaluation of the ICIAR 2018 Breast Cancer Histology (BACH) dataset demonstrates that LExNet achieves a superior validation accuracy of approximately 98.3%, surpassing traditional deep-learning models by approximately 12% while reducing computational demands by 40-50%. The autoencoder-driven feature extraction significantly improved noise and dimensionality reduction, effectively decreasing overfitting risk by over 30%. Additionally, PSO-driven hyperparameter tuning notably accelerated convergence and reduced manual hyperparameter tuning time by up to 70%. Grad-CAM interpretability analysis further confirms that LExNet's predictions closely align with critical pathological features recognized by expert pathologists, highlighting its clinical relevance. CONCLUSION: Consequently, LExNet offers a robust, interpretable, and computationally efficient solution for real-time breast cancer diagnostics, making it particularly suitable for resource-constrained clinical settings.