Breast cancer classification using breast ultrasound images with a hybrid of transfer learning and Bayesian-optimized fast learning network

Abstract Breast cancer is the most prevalent cancer among women globally, with over 2.3 million new cases reported annually, hence, early and accurate diagnosis is crucial in minimizing mortality rates. In light of the limitation on the conventional interpretation of images, this study presents a novel hybrid model that integrates DenseNet201-based transfer learning with a Bayesian-Optimized Fast Learning Network (FLN) for breast cancer classification from ultrasound images. DenseNet201 was employed to obtain robust, high-quality features from pre-trained weights. FLN is finely tuned with Bayesian optimization to select optimal hyperparameters such as learning rate, hidden neurons, and dropout rate. The proposed model achieved an accuracy of 96.79%, an F1 score of 94.71%, a precision of 96.81%, and a recall of 93.48% with AUC scores of 0.96, 0.95, and 0.98 for benign, malignant, and normal classes. These results underscore the model’s balanced performance and its ability to minimize misclassifications, particularly false positives. The end-to-end hybrid approach not only outperforms several state-of-the-art models but also demonstrates improved generalization and stability, offering a promising, clinically viable tool for enhancing breast cancer diagnosis.