Early detection and reliable diagnosis of breast cancer is vital for improving the greatness of life of patients and clinical decision making. Here we propose a deep learning approach based on a Multilayer Perceptron (MLP) model for binary classification of malignant and benign tissue samples. We preprocessed the Kaggle breast cancer dataset, through feature standardization and label encoding, and stratified train-test split to avoid learning imbalance. We trained an MLP with three hidden layers with dropout regularization, optimised using the Adam algorithm, using binary cross-entropy loss and EarlyStopping to avoid overfitting. The test accuracy of the proposed model, both when used alone and in conjunction with other models, was calculated to be 98.83%, while the AUC was found to be 0.9968, with great discriminative capacity. The confusion matrix analysis confirmed the high reliability of the classification, since among 61 malignant cases, only 2 were misclassified as benign and in contrast to that, there are no misdiagnoses in the case of false positives. Furthermore, SHAP explainability assessment pointed to concave points_mean, concave points_worst, and concavity_worst as the leading predictors of malignant behavior in concordance with clinical tumor morphology metrics. Diagnostic accuracy of the developed MLP model was 98.86±0.52%, which is high, and the interpretability of decision making through feature contributions values are validated. Thus, this framework holds promise for integration into computer-aided diagnosis systems for aiding medical practitioners in the early detection of breast cancer.

Keywords: Breast Cancer Detection; Deep Learning; Explainable Artificial Intelligence (XAI); SHAP Interpretation