BreastCNet: Breast Cancer Detection, Classification, and Localization Convolutional Neural Network With Advanced Optimization Techniques

Breast cancer remains a significant global health issue, primarily affecting females and requiring advanced detection methods to improve patient outcomes. While computer-aided diagnostic methods have progressed, limitations in accuracy persist. This study introduced BreastCNet, a Convolutional Neural Network (CNN) with hyperparameter optimization and a multi-task learning framework, enhancing classification and lesion localization. The BreastCNet integrated CNNs with the Grey Wolf Optimizer (GWO) and Parrot Optimizer (PO) in a dual-optimization approach. GWO fine-tuned neuron counts in dense layers to enhance feature learning, while PO dynamically adjusted the learning rate to improve convergence. Specifically, the parrot optimizer adjusted the learning rate from 0.001 to 0.00156, resulting in a 1.5% improvement in accuracy, and the grey wolf optimizer fine-tuned neuron numbers (Dense_1: 6, Dense_2: 866), contributing to a 2.3% increase in validation accuracy. These optimizations mitigated overfitting, accelerated convergence, and improved generalization, making the BreastCNet computationally efficient and clinically viable. The multi-task learning framework simultaneously performed breast ultrasound image classification (benign/malignant) and lesion localization via bounding box regression. The model achieved 98.10% validation accuracy, an AUC of 0.995, an F1-score of 0.98, and an IoU score of 0.96 for lesion localization on the Breast Ultrasound Images (BUSI) dataset, with comparable improvements on the BUSI, Digital Database for Screening Mammography (DDSM), and INbreast datasets. The BreastCNet is highly likely to advance breast cancer detection and lesion localization, providing precise localization and improved generalization, making it a promising tool for clinical applications.