Breast cancer (BC) is a major global health challenge, requiring efficient computational systems for early detection and precise diagnosis. This paper presents DNN-based classification framework for breast cancer detection evaluated using a stratified data-splitting. Rather than proposing a new network architecture, the framework aims to improve model stability and class balance in breast cancer classification using the Wisconsin Diagnostic Breast Cancer (WDBC) dataset. The dataset was divided into training, validation, and testing subsets (64%–16%–20%) to ensure fair model evaluation and prevent overfitting. Instead of standardization, normalization was applied to enhance learning efficiency. Experimental results demonstrate an accuracy of 98.25%, precision of 98.61%, and recall of 98.61%, indicating a high level of reliability in distinguishing malignant from benign cases. The ROC–AUC score of 0.98 confirms the model’s excellent discriminative ability. Compared with recent studies, the final DNN model achieves competitive performance. The findings highlight the importance of rigorous evaluation protocols when assessing DNN-based breast cancer classification models.