Related Experiment Video
Updated: Jun 8, 2025
13:44
Detection of Architectural Distortion in Prior Mammograms via Analysis of Oriented Patterns
Published on: August 30, 2013
42.8K
Combination of Deep Learning Grad-CAM and Radiomics for Automatic Localization and Diagnosis of Architectural
Xiao Chen1, Yang Zhang2, Jiejie Zhou3
1Department of Radiology, First Affiliated Hospital of Wenzhou Medical University, Wenzhou, China (X.C., J.Z., Y.P., Y.S., G.C., M.W.).
Academic Radiology
|November 4, 2024
Summary
Artificial intelligence using deep learning can detect architectural distortion on digital breast tomosynthesis, with performance comparable to manual analysis. This AI approach shows promise for improving breast cancer diagnosis.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Radiology
Background:
- Architectural distortion (AD) detection on digital breast tomosynthesis (DBT) presents diagnostic challenges.
- Artificial intelligence (AI) and deep learning (DL) offer potential solutions for improving diagnostic accuracy.
Purpose of the Study:
- To apply AI with deep learning algorithms for detecting architectural distortion (AD) on DBT.
- To utilize radiomics for classifying AD and estimating malignancy probability.
- To compare the diagnostic performance of AI-generated regions of interest (ROIs) with manually delineated ROIs.
Main Methods:
- A dataset of 500 cases with AD on DBT reports was used, with 292 for training and 208 for testing.
- Deep learning (DL) with Gradient-weighted Class Activation Mapping (Grad-CAM) was employed for automated abnormality localization and ROI generation.
- Radiomics analysis was performed on both DL-generated and manually delineated ROIs to estimate malignancy probability.
- Cases were categorized into pure AD and AD associated with other features (mass, high-density, calcifications).
Main Results:
- The overall malignancy rate was 57%. Pure AD had a significantly lower malignancy rate (39.7%) compared to AD associated with other features (76.8%).
- The area under the curve (AUC) for diagnostic performance was 0.82 with manual ROI and 0.84 with DL-generated ROI in the testing set.
- For pure AD cases, DL-generated ROI achieved an AUC of 0.77, while AD associated with other features yielded an AUC of 0.86.
Conclusions:
- Deep learning (DL) effectively detects architectural distortion (AD) on digital breast tomosynthesis (DBT), with diagnostic performance comparable to manual ROI delineation.
- The AI strategy demonstrated effectiveness for pure AD detection, although performance was superior for AD cases associated with other mammographic features.
