Enhancing Breast Cancer Classification Accuracy Using Data Augmentation And CNN

Globally, breast cancer consistently ranks as a leading cancer diagnosis, underscoring the urgency for refined diagnostic methodologies. In this exploration, we harness the capabilities of Convolutional Neural Networks (CNNs), further amplified by data augmentation strategies, to elevate the precision in classifying histopathological images indicative of breast cancer. Anchoring our research is the BreakHis dataset, a collection of 9,109 microscopic breast tumor images. Against this backdrop, we conducted a rigorous assessment of five prominent CNN models: InceptionV3 V1, DenseNet201, ResNet50, hybrid model, and VGG16. Our research underscores the pivotal role of data augmentation in addressing dataset limitations, such as class imbalance and overfitting. By introducing controlled variations in the training images, we observed a marked improvement in model generalization across all architectures. Among the models, VGG16 and DenseNet201 emerged as frontrunners, achieving accuracies of 91% and 90%, respectively. Notably, while the hybrid model exhibited an impressive training accuracy of 99.83%, its validation accuracy was limited to 69.53%, hinting at potential overfitting. Furthermore, our comparative analysis highlighted each model's distinct strengths and weaknesses, offering insights into their applicability in real-world diagnostic scenarios. The precision and recall metrics provided a deeper understanding of each model's capability to differentiate between benign and malignant samples, with VGG16 demonstrating superior performance. This study accentuates the synergistic potential of CNNs and data augmentation in advancing breast cancer diagnostics. Our findings pave the way for further research, emphasizing the need for tailored model architectures and augmentation strategies to harness the full potential of deep learning in oncology.