Related Experiment Video
Updated: Jun 27, 2026
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
Explainable Breast Cancer Detection Using Hierarchical Multi-Scale and Edge-Aware Transformer Networks
Maria Altaib Badawi1, Ehtisham Arshad2, Armughan Ali3
1Department of Computer Science and Information, College of Science Zulfi, Majmaah University, Al-Majmaah 11952, Saudi Arabia.
Bioengineering (Basel, Switzerland)
|June 26, 2026
Summary
This study introduces an efficient deep learning framework for accurate breast cancer classification from mammograms. The model achieves over 99% accuracy, enhancing early detection and improving upon existing methods with explainable AI.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Breast cancer is a leading cause of death in women, necessitating improved early detection methods.
- Mammography is crucial for early detection, but manual interpretation faces challenges due to image volume and subtle lesion variations.
- Current deep learning models for breast cancer classification struggle with overfitting, complexity, generalization, and interpretability.
Purpose of the Study:
- To propose a computationally efficient, context-aware deep learning framework for breast cancer classification.
- To integrate multi-scale attention mechanisms and explainable AI (XAI) for enhanced diagnostic performance and interpretability.
- To address limitations of existing models, including overfitting, computational cost, and generalization.
Main Methods:
- Developed a framework integrating Hierarchical Multi-Scale Transformer (HMT) and Edge-Aware Local Transformer (ELT) modules for global and local feature extraction.
- Incorporated an Adaptive Contextual Refinement (ACR) module to maintain feature consistency across resolutions.
- Utilized a Meta-Ensemble Classification (MEC) framework with weighted SVM and K-Nearest Neighbors (KNN) classifiers.
- Employed XAI techniques including Grad-CAM, SHAP, Occlusion Sensitivity Analysis (OSA), and TIxAI for explainability.
Main Results:
- Achieved superior accuracy exceeding 99% across four benchmark mammography datasets (CBIS-DDSM, DDSM, INBreast, MIAS).
- Outperformed transformer baselines (Swin-T, ViT) with lower parameter complexity and higher accuracy on CBIS-DDSM.
- Demonstrated strong cross-dataset generalization with high precision, recall, and F1-scores.
- Explainability analysis showed consistent saliency maps, though external clinical validation is pending.
Conclusions:
- The proposed framework offers an effective and computationally efficient solution for automated breast cancer classification.
- The integration of multi-scale attention and XAI significantly improves diagnostic accuracy and model interpretability.
- The model shows promise for clinical application in improving breast cancer detection rates and patient outcomes.