BAG-DNN: A Novel Attention-Gated Deep Neural Network Ensemble for Interpretable Breast Cancer Diagnosis
Abstract
Breast cancer remains one of the leading causes of cancer mortality among women worldwide, and computer-aided diagnosis (CAD) systems built on machine learning have become an important second reader for radiologists and pathologists. While deep neural networks (DNNs) can match or exceed classical classifiers on many tasks, their black-box nature limits clinical trust, and on small tabular clinical datasets they do not always outperform simpler, well-regularized models. This paper proposes an Attention-Gated Deep Neural Network (AG-DNN) for tabular breast-cancer diagnosis, in which a learnable, sigmoid-based feature-wise gate re-weights the 30 input clinical measurements before they enter a standard dense network. The gate is centered so that the model is initialized identically to a plain DNN and only learns to emphasize or suppress features where doing so improves the objective, guaranteeing that the attention mechanism can never reduce representational capacity relative to the baseline. Five independently initialized AG-DNN members are combined by probability averaging into a Bagged AG-DNN (BAG-DNN), the final proposed model. Using the Wisconsin Diagnostic Breast Cancer (WDBC) dataset (569 patients, 30 features) and stratified 5-fold cross-validation, BAG-DNN is compared against Logistic Regression, Random Forest, an RBF Support Vector Machine, a scikit-learn multilayer perceptron, a single plain DNN, and a plain-DNN ensemble. BAG-DNN improves mean cross-validated accuracy from 0.9666 (single AG-DNN) to 0.9702 and reduces the standard deviation of AUC from 0.0085 to 0.0040, reaching a mean ROC-AUC of 0.9954 — statistically indistinguishable from the strongest baselines (Logistic Regression 0.9953, MLP 0.9959) — while remaining fully interpretable: the learned attention weights consistently up-weight the ‘worst’ (most extreme) tumor-morphology measurements, in agreement with established clinical and statistical findings about this dataset. We report all results transparently, including cases where the attention mechanism does not exceed a simple ensembled plain DNN on raw accuracy, and discuss this trade-off between interpretability and accuracy gain.