Related Experiment Video
Updated: Oct 10, 2025
13:44
Detection of Architectural Distortion in Prior Mammograms via Analysis of Oriented Patterns
Published on: August 30, 2013
43.1K
A convolutional deep learning model for improving mammographic breast-microcalcification diagnosis
Daesung Kang1, Hye Mi Gweon2, Na Lae Eun2
1Department of Healthcare Information Technology, Inje University, Gimhae, Republic of Korea.
Scientific Reports
|December 15, 2021
Summary
Deep convolutional neural networks (DCNNs) show promise in classifying breast microcalcifications from mammograms. The ResNet-101 model demonstrated high accuracy and specificity, potentially reducing unnecessary biopsies.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Breast microcalcifications are key indicators of breast cancer on screening mammograms.
- Accurate classification of microcalcifications as malignant or benign is crucial for patient management and reducing unnecessary invasive procedures.
Purpose of the Study:
- To evaluate the diagnostic performance of deep convolutional neural networks (DCNNs) for classifying breast microcalcifications.
- To compare the effectiveness of several pre-trained DCNN models and an ensemble model in this classification task.
Main Methods:
- Retrospective analysis of 1579 mammographic images with suspicious microcalcifications.
- Utilized five pre-trained DCNN models (trained on ImageNet) and an ensemble model for classification.
- Employed Gradient-weighted Class Activation Mapping (Grad-CAM) to validate model focus on microcalcification regions.
Main Results:
- The ensemble model achieved the highest Area Under the Curve (AUC) of 0.856.
- DenseNet-201 model showed the best sensitivity (82.47%) and Negative Predictive Value (NPV) (86.92%).
- ResNet-101 model achieved the highest accuracy (81.54%), specificity (91.41%), and Positive Predictive Value (PPV) (81.82%).
Conclusions:
- DCNNs, particularly ResNet-101, exhibit significant effectiveness in diagnosing breast microcalcifications.
- The high specificity and PPV of ResNet-101 suggest a potential to decrease the rate of unnecessary breast biopsies.
