Related Experiment Video
Updated: Jun 11, 2025
13:44
Detection of Architectural Distortion in Prior Mammograms via Analysis of Oriented Patterns
Published on: August 30, 2013
42.8K
Optimizing mammography interpretation education: leveraging deep learning for cohort-specific error detection to
Xuetong Tao1, Warren M Reed1, Tong Li2,3
1The University of Sydney, Faculty of Health Sciences, Discipline of Medical Imaging Science, Sydney, New South Wales, Australia.
Journal of Medical Imaging (Bellingham, Wash.)
|October 7, 2024
Summary
This study shows convolutional neural networks (CNNs) can identify difficult false-positive (FP) regions in mammograms. This approach can tailor training for radiologists, improving accuracy in mammogram interpretation.
Area of Science:
- Radiology
- Artificial Intelligence
- Medical Imaging
Background:
- Accurate mammogram interpretation is crucial but challenging.
- False-positive (FP) errors in mammography can lead to unnecessary patient anxiety and procedures.
- Personalized training strategies for radiologists show promise in reducing interpretation errors.
Purpose of the Study:
- To investigate the feasibility of using convolutional neural networks (CNNs) with transfer learning.
- To categorize regions of false-positive (FP) errors in screening mammograms.
- To predict the likelihood of FP detections based on radiologist cohorts with similar geographic characteristics.
Main Methods:
- Utilized mammography test sets from two geographically distinct radiologist cohorts.
- Segmented FP patches and categorized them as "difficult" or "easy" based on reader consensus.
- Developed a patch-wise CNN model using ResNet-50 for feature extraction and binary classification, assessed via 10-fold cross-validation.
Main Results:
- The transferred ResNet-50 architecture achieved high performance.
- Achieved receiver operating characteristics area under the curve (AUC) values of 0.933 for cohort A and 0.975 for cohort B.
- Demonstrated superior performance compared to other evaluated architectures.
Conclusions:
- CNN-based transfer learning is feasible for predicting the difficulty of FP patches in mammograms.
- This approach can be tailored to specific radiologist cohorts based on geographic characteristics.
- Highlights a potential method for personalized mammography training to reduce errors.
