Enhancing Diagnostic Precision in Breast Cancer Classification Through EfficientNetB7 Using Advanced Image Augmentation and Interpretation Techniques

Abstract

The precise classification of breast ultrasound images into benign, malignant, and normal categories represents a critical challenge in medical diagnostics, exacerbated by subtle interclass variations and the variable quality of clinical imaging. State-of-the-art approaches largely capitalize on the advanced capabilities of deep convolutional neural networks (CNNs), with significant emphasis on exploiting architectures like EfficientNet that are pre-trained on extensive datasets. While these methods demonstrate potential, they frequently suffer from overfitting, reduced resilience to image distortions such as noise and artifacts, and the presence of pronounced class imbalances in training data. To address these issues, this study introduces an optimized framework using the EfficientNetB7 architecture, enhanced by a targeted augmentation strategy. This strategy employs aggressive random rotations, color jittering, and horizontal flipping to specifically bolster the representation of minority classes, thereby improving model robustness and generalizability. Additionally, this approach integrates an adaptive learning rate scheduler and implements strategic early stopping to refine the training process and prevent overfitting. This optimized model demonstrates a substantial improvement in diagnostic accuracy, achieving a 98.29% accuracy rate on a meticulously assembled test dataset. This performance significantly surpasses existing benchmarks in the field, highlighting the model's enhanced ability to navigate the intricacies of breast ultrasound image analysis. The high diagnostic accuracy of this model positions it as an invaluable tool in the early detection and informed management of breast cancer, potentially transforming current paradigms in oncological care.