This study is concerned with optimizing performance on the use of ML algorithms in breast cancer forecast, with paramount concern over the data of medical images. Leverage the power of convolutional neural networks; the approach is designed to correctly classify benign and malignant breast masses. The study can be divided into two components: exploration of model training dynamics and the assessment of performance metrics. It will first consider the accuracy and loss against epochs curve to understand the model’s learning behavior. A positive correlation between epochs and accuracy has been noted, while the onset of overfitting is often emphasized beyond a specific number of epochs. Some significant mitigation strategies that include regularization techniques alongside data augmentation are discussed in length. For the second half of the paper, a more exhaustive analysis about the metric used for performance evaluation is presented. These contain accuracy, precision, recall, F1-score, ROC-AUC score, and Cohen Kappa score. It is evident that the model is able to depict commendable accuracy and precision, but it has further scope for improvement in recalling malignant cases and their F1-score. Hence, a general conclusion could be drawn as such that there is still a significant necessity to further optimize the models and evolve them in making them more strong and sound for breast cancer prediction. The article would aim at contributing to the overcoming of the limitations of breast cancer diagnosis in standings of sensitivity besides specificity, mainly by addressing overfitting, and then optimizing related model performance metrics.