A Deep Learning-Based Artificial Neural Network for Breast Cancer Classification: A Comprehensive Performance Evaluation

Abstract

Accurate and prompt diagnosis of cancers is critical to improve patient outcomes and to make timely clinical decisions. The present study built a multilayer feedforward artificial neural network (ANN) deep learning model to identify between malignant and healthy instances. The model included 186 inputs, one output neuron and four hidden layers with 64, 32, 16 and 8 neurons each. The network was trained with Scaled Conjugate Gradient (SCG) technique. The cross-entropy loss function was minimized and early stopping was applied on validation performance. This strategy was designed to improve generality and reduce over-fitting. The developed model achieved an F1-score of 93.07%, an Area under the Receiver Operating Characteristic Curve (AUC) of 0.8045, and an overall classification accuracy of 87.07 percent. Together these metrics give a significant discrimination between malignant and normal samples. The model was validated using the Receiver Operating Characteristic (ROC) analysis where it was found that the model was confident in its categorization as the cancer patients always received a higher prediction probability than healthy instances. The training was successful and exhibited steady learning, converging in 69 epochs with a final cross-entropy loss of 0.295. The confusion matrix revealed that the model was highly sensitive to correctly identify most of the cancer cases. The considerably reduced specificity was mostly due to the considerable imbalance in the sample class. The learning curves were smooth and did not indicate any overfitting. Further error analysis indicated that performance was generally age-independent, and there was no appreciable drop in accuracy with age. To further illustrate the robustness of the neural network, a comprehensive performance dashboard was deployed to consolidate many evaluation measures, dataset characteristics, prediction distributions, and optimization behavior into a single assessment. The proposed artificial neural network provides a reliable framework with strong classification performance and feasible computational features for cancer prediction in general. Future studies should aim on enhancing the specificity, e.g. via enhanced feature selection, threshold optimization and class balance. The results to-date highlights the promise of deep learning for medical diagnostics and cancer risk prediction. The study highlights the need to apply the latest machine learning techniques to improve the quality of patient care and accuracy of cancer diagnoses.