Related Experiment Video
Updated: Jul 27, 2025
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
1.9K
Transformer-based Deep Neural Network for Breast Cancer Classification on Digital Breast Tomosynthesis Images
Weonsuk Lee1, Hyeonsoo Lee1, Hyunjae Lee1
1From Lunit, 5F, 374 Gangnam-daero, Gangnam-gu, Seoul 06241, Republic of Korea.
Radiology. Artificial Intelligence
|June 9, 2023
Summary
A new transformer-based deep neural network improves breast cancer detection on digital breast tomosynthesis (DBT) images by analyzing neighboring sections. This efficient model outperforms traditional methods and enhances diagnostic accuracy.
Area of Science:
- Radiology and Medical Imaging
- Artificial Intelligence in Healthcare
- Oncology
Background:
- Digital breast tomosynthesis (DBT) is an advanced imaging technique for breast cancer screening.
- Accurate detection of breast cancer in DBT images remains a challenge, necessitating improved computational methods.
- Deep learning models offer potential for enhancing diagnostic performance in medical imaging analysis.
Purpose of the Study:
- To develop an efficient deep neural network (DNN) model for breast cancer detection in DBT images.
- To leverage contextual information from neighboring image sections for improved detection accuracy.
- To compare the performance of the proposed DNN model against existing architectures.
Main Methods:
- A transformer-based deep neural network architecture was employed to analyze contextual information across DBT image sections.
- The proposed model was trained and validated on a large dataset of 5174 four-view DBT studies.
- Performance was evaluated against two baseline models (3D convolutions and 2D per-section analysis) using metrics like AUC, sensitivity, and specificity.
Main Results:
- The transformer-based model demonstrated significantly improved classification performance compared to the per-section baseline.
- Key performance metrics showed substantial gains: AUC increased from 0.88 to 0.91, sensitivity from 81.0% to 87.7%, and specificity from 80.5% to 86.4%.
- The transformer model achieved comparable classification accuracy to a 3D convolution model while utilizing only 25% of the computational resources.
Conclusions:
- A transformer-based DNN effectively utilizes contextual data from neighboring DBT sections to enhance breast cancer detection.
- The proposed model offers superior diagnostic performance and computational efficiency compared to existing methods.
- This approach represents a significant advancement in AI-driven breast cancer diagnosis using DBT imaging.
