An improved XAI-based DenseNet model for breast cancer detection using reconstruction and fine-tuning

Breast cancer remains a major public health concern and a leading cause of cancer-related deaths among women worldwide. Early and accurate diagnosis is crucial for improving patient outcomes and reducing mortality. This study proposes a novel Explainable AI (XAI) based deep-learning (DL) approach that enhances breast cancer detection by integrating advanced image preprocessing, DenseNet architectural modifications and fine-tuning process tailored for histopathological imaging. The key novelty of this work lies in the strategic enhancement of DenseNet with BN-ReLU-Conv and Block-End layers, along with optimized fine-tuning techniques, which improve feature extraction and classification accuracy. The proposed approach is evaluated on five benchmark histopathology datasets: BreakHis 40X, 100X, 200X, 400X, and BACH. Experimental results demonstrate that DenseNet169 outperforms other models, achieving remarkable accuracy scores of 99.50%, 98.80%, 97.27%, and 96.98% for BreakHis 40X, 100X, 200X, 400X, and 94.75% for the BACH dataset, with minimal false positives. To enhance model transparency and clinical trust, we incorporate XAI techniques such as Class Activation Maps (CAM) and Saliency Maps, along with their combined visualization, to provide intuitive and interpretable explanations of the model's decision-making process. These techniques help highlight the most relevant regions contributing to cancer detection, offering valuable insights for pathologists and clinicians. Despite its high performance, the approach has certain limitations, while the model performs well in controlled environments, we did not develop a clinical device for real-world deployment. Future work will focus on incorporating these techniques and translating the model into clinical practice for more effective breast cancer detection.