Abstract
Breast cancer diagnosis using deep learning models often suffers from limited interpretability, unstable attention reasoning, and inadequate transparency in lesion-sensitive decision processes, thereby restricting clinical trust and deployment. This study proposes ASYMIRX, a novel asymmetry-guided explainable deep learning framework designed to improve interpretable breast cancer image classification through transparent feature reasoning, counterfactual sensitivity analysis, uncertainty quantification, and latent-space interpretability. Unlike conventional black-box explainable AI approaches, the proposed framework integrates asymmetry-aware attention propagation, Grad-CAM reasoning cascades, counterfactual malignancy tracing, and uncertainty decomposition to reveal clinically meaningful feature evolution across hierarchical network layers. The framework combines convolutional feature extraction, semantic attention reweighting, latent representation analysis, and Monte Carlo uncertainty estimation to evaluate lesion morphology, border irregularity, tissue asymmetry, and contextual patterns. Experimental results demonstrated substantial improvements over the baseline Mirai model, achieving a 42.6% improvement in explanation faithfulness, a 22.5% improvement in localization accuracy, a 17.7% enhancement in attribution stability, a 56.3% improvement in latent cluster compactness, and a 27.7% reduction in predictive uncertainty entropy. Furthermore, ASYMIRX generated more stable semantic attention distributions and clinically coherent lesion localization, demonstrating superior interpretability, robustness, and diagnostic reliability suitable for trustworthy medical AI applications.
| Original language | English |
|---|---|
| Article number | 100274 |
| Number of pages | 21 |
| Journal | Computer Methods and Programs in Biomedicine Update |
| Volume | 10 |
| Early online date | 21 Oct 2026 |
| DOIs | |
| Publication status | E-pub ahead of print - 21 Oct 2026 |
UN SDGs
This output contributes to the following UN Sustainable Development Goals (SDGs)
-
SDG 3 Good Health and Well-being
Fingerprint
Dive into the research topics of 'ASYMIRX: Asymirai-Guided Explainable Deep Learning for Interpretable and Trustworthy Breast Cancer Image Classification'. Together they form a unique fingerprint.Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver