Explainable Deep Learning for Breast Cancer Diagnosis Using NODE and FT-Transformer Models: A SHAP and LIME-Based Interpretability Study

Early and accurate detection of breast cancer is vital for improving patient outcomes and recent advancements in deep learning models offer promising solutions for assisting in diagnostic decision-making. In this study, classification process of breast cancer is investigated by utilizing two advanced deep learning architectures particularly developed for tabular structured datasets, Neural Oblivious Decision Ensembles (NODE) and the FT-Transformer. Additionally, for ensuring transparent and trustworthy model behavior, two explainability techniques, SHAP (SHapley Additive exPlanations) and LIME (Local Interpretable Model-agnostic Explanations) are applied to interpret the influence of individual features on model estimations. Experimental results demonstrate that both models achieved high classification performance, with FT-Transformer slightly outperforming NODE in terms of general performance. Moreover, explainability analyses consistently indicate Bare Nuclei feature as the most decisive feature, while also revealing differences in feature contributions between the utilized models. Consequently, the findings in this study emphasize the different prediction mechanisms of both models and analyze the importance of integrating explainable artificial intelligence (AI) techniques into clinical prediction tasks.