Design of a PSO-Optimized VGG19 Convolutional Neural Network for Automatic Breast Cancer Image Classification

Background: Human Epidermal Growth Factor Receptor 2 (HER2) overexpression is detected in approximately 25–30% of malignant breast cancer cases and is intrinsically associated with aggressive tumor progression. Reliable assessment of HER2 status is critical for selecting optimal targeted therapies; however, conventional immunohistochemistry (HER2 SISH) evaluation predominantly relies on visual inspection, a process that is highly subjective and susceptible to inter-observer variability. To mitigate diagnostic inconsistencies, this study proposes an automated classification framework leveraging a VGG19 convolutional neural network, fundamentally enhanced via Particle Swarm Optimization (PSO) for rigorous hyperparameter tuning. The model development and evaluation in this study were performed using a purely synthetic image dataset. Methods: The PSO algorithm is systematically employed to optimize critical training parameters, facilitating an efficient exploration of the complex search space and significantly improving the stability of the learning process. Experimental evaluations conducted on these purely synthetic high-resolution histopathology images demonstrate that the proposed PSO-VGG19 framework substantially outperforms baseline architectures, including MobileNet. Results: Specifically, the optimized system achieves a classification accuracy of 98.10%, sensitivity of 98.11%, specificity of 99.37%, and an Area Under the Curve (AUC) of 99.66%. Conclusions: These empirical results demonstrate that integrating metaheuristic optimization with deep learning architectures yields robust computational performance for HER2 patch-level classification, serving as a foundational computational proof-of-concept for future digital pathology research.