U-Net Models for Breast Cancer Detection: Improving Diagnostic Accuracy and Specificity

Breast cancer remains a critical global health issue, necessitating continuous research and innovative approaches for diagnosis, treatment, and prevention. This study evaluates the effectiveness of U -Net models in enhancing diagnostic precision and efficiency using real hospital samples. We aim to improve key diagnostic metrics such as accuracy, sensitivity, and specificity through the application of U-Net models. Our image classification model, tailored for $256 \times 256 \times 3$ input images, excels in detecting and categorizing tumor cells. The architecture begins with initial convolutional layers featuring 64 filters, progresses to layers with 128 filters, and includes a Dropout layer to prevent overfitting. The deep network for object detection utilizes both region proposal and regression/classification approaches, achieving $\mathbf{9 2. 2 7 \%}$ confidence and $100 \%$ accuracy. Additionally, our deep learning algorithms accurately segment nuclei in histopathological images, employing a clustering strategy that delivers $88.81 \%$ confidence and $100 \%$ accuracy. Visual results demonstrate precise tumor cell localization and prediction confidence. Performance metrics from ten experimental runs indicate average confidence levels between $\mathbf{7 4. 1 9 \%}$ and $\mathbf{9 2. 3 1 \%}$, with $\mathbf{9 0. 0 \%}$ accuracy and specificity in benign analysis. The model’s ability to classify non-carcinomas versus carcinomas achieved an AUC of 0.78, illustrating its effective differentiation between classes.