Related Experiment Videos
An attention-based transfer learning framework for breast cancer classification in mammography under limited data
1Department of Electronics and Communication Engineering, National Institute of Technology Hamirpur, Hamirpur, Himachal Pradesh, India.
Frontiers in Medical Technology
|July 24, 2026
Summary
This study enhances breast cancer detection using deep learning models, MobileNetV2-SE and ResNet50-SE, achieving high accuracy on mammography images. These AI frameworks improve early diagnosis and aid computer-aided diagnosis for better patient outcomes.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Breast cancer is a significant global health concern, necessitating improved diagnostic tools for early detection and increased survival rates.
- Current diagnostic methods for breast cancer require enhancement for accuracy and efficiency.
- Computer-aided diagnosis (CAD) systems show promise in improving mammogram interpretation.
Purpose of the Study:
- To evaluate the efficacy of transfer learning-based deep learning models, specifically MobileNetV2 and ResNet50, for binary breast cancer classification.
- To investigate the impact of the Squeeze-and-Excitation (SE) attention mechanism on model performance for breast cancer detection.
- To assess the diagnostic accuracy of proposed models on diverse mammography datasets.
Main Methods:
- Utilized transfer learning with pre-trained MobileNetV2 and ResNet50 architectures.
- Integrated the Squeeze-and-Excitation (SE) attention mechanism to enhance feature learning.
- Performed partial model fine-tuning by freezing early layers.
- Applied Gradient-weighted Class Activation Mapping (Grad-CAM) for model interpretability.
- Evaluated performance on CLAHE-binary, DDSM, INbreast, and MIAS datasets using stratified k-fold cross-validation.
Main Results:
- The MobileNetV2-SE model achieved high mean accuracies: 99.14% (DDSM), 93.53% (INbreast), 92.35% (MIAS), and 96.33% (CLAHE-binary).
- The ResNet50-SE model demonstrated strong performance with mean accuracies: 100% (DDSM), 88.89% (INbreast), 71.60% (MIAS), and 97.98% (CLAHE-binary).
- Both models exhibited reliable and competitive performance, improving diagnostic accuracy in mammogram classification.
Conclusions:
- The developed deep learning frameworks, particularly MobileNetV2-SE, offer effective and accurate solutions for automated breast cancer detection from mammograms.
- The integration of SE attention mechanisms and partial fine-tuning enhances the diagnostic capabilities of deep learning models in medical imaging.
- These findings contribute to advancing computer-aided diagnosis systems for breast cancer, supporting radiologists and improving patient care.