Related Experiment Video
Updated: Nov 5, 2025
13:44
Detection of Architectural Distortion in Prior Mammograms via Analysis of Oriented Patterns
Published on: August 30, 2013
43.2K
A generative adversarial network-based abnormality detection using only normal images for model training with
Albert Swiecicki1, Nicholas Konz2, Mateusz Buda3
1Department of Electrical and Computer Engineering, Duke University, Durham, NC, USA. albert.swiecicki@duke.edu.
Scientific Reports
|May 14, 2021
Summary
This study introduces a novel deep learning method for cancer detection in medical images. By training an inpainting network on healthy images, it identifies abnormalities through completion errors, improving cancer screening accuracy.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Deep Learning
Background:
- Object detection in images is challenged by limited data, especially in cancer screening like digital breast tomosynthesis (DBT).
- Less than 1% of DBT cases contain cancer, posing a significant data scarcity problem for training detection models.
Purpose of the Study:
- To develop a deep learning method for cancer detection using only images without cancer.
- To leverage generative adversarial networks (GANs) for identifying abnormalities in medical scans.
Main Methods:
- Trained an inpainting generative adversarial network (GAN) using a dataset of normal digital breast tomosynthesis (DBT) images.
- During inference, parts of images were removed, and the GAN attempted to inpaint them.
- Abnormal locations were identified by significant errors in the GAN's image completion process.
Main Results:
- Experiments on a large DBT dataset (19,230 volumes) showed higher completion errors in cancer locations compared to non-cancer locations.
- The mean error ratio for cancer locations was 2.77, indicating a significant difference.
- The method successfully distinguished between cancerous and non-cancerous regions based on image completion accuracy.
Conclusions:
- The proposed inpainting GAN method is effective for cancer detection in DBT, even with limited positive samples.
- This approach offers a promising solution for improving cancer screening in data-scarce scenarios.
- Publicly available data from Duke University was utilized, promoting research reproducibility.
