Related Experiment Video
Updated: Mar 11, 2026
15:48
Tracking the Mammary Architectural Features and Detecting Breast Cancer with Magnetic Resonance Diffusion Tensor Imaging
Published on: December 15, 2014
23.4K
Mass detection in digital breast tomosynthesis: Deep convolutional neural network with transfer learning from
Ravi K Samala1, Heang-Ping Chan1, Lubomir Hadjiiski1
1Department of Radiology, University of Michigan, Ann Arbor, Michigan 48109.
Medical Physics
|December 3, 2016
Summary
This study developed a deep convolutional neural network (DCNN) for detecting masses in digital breast tomosynthesis (DBT) using transfer learning from mammograms, achieving 91% sensitivity at 1 false positive per DBT volume.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiology
Background:
- Digital breast tomosynthesis (DBT) is an advanced mammography technique.
- Computer-aided detection (CAD) systems aim to improve mass detection accuracy.
- Developing CAD systems for DBT presents challenges due to data requirements.
Purpose of the Study:
- To develop a CAD system for mass detection in DBT volumes.
- To utilize a deep convolutional neural network (DCNN) with transfer learning from mammograms.
- To compare the DCNN-based CAD system with a previously developed feature-based CAD system.
Main Methods:
- A DCNN was trained on mammographic images and then fine-tuned using DBT data via transfer learning.
- The DCNN model incorporated data augmentation, background correction, jittering, and dropout techniques.
- Performance was evaluated using free-response ROC curves, comparing sensitivity and false positive rates.
Main Results:
- Transfer learning improved DCNN performance on DBT masses from an AUC of 0.81 to 0.90.
- The DCNN-based CAD system achieved 91% sensitivity at 1 FP/DBT volume, outperforming the feature-based system (83% sensitivity).
- The performance difference between the two CAD systems was statistically significant (p < 0.05).
Conclusions:
- Image patterns learned from mammograms can be effectively transferred to DBT mass detection using DCNNs.
- Large mammography datasets are valuable for developing new CAD systems for DBT.
- This approach alleviates the need for extensive new data collection for emerging imaging modalities.