Related Experiment Video
Updated: Jul 16, 2025
13:44
Detection of Architectural Distortion in Prior Mammograms via Analysis of Oriented Patterns
Published on: August 30, 2013
42.9K
FSE-Net: feature selection and enhancement network for mammogram classification.
Caiqing Liao1, Xin Wen1, Shuman Qi1
1College of Software Engineering, Taiyuan University of Technology, Taiyuan 030600, People's Republic of China.
Physics in Medicine and Biology
|September 15, 2023
Summary
A new Feature Selection and Enhancement Network (FSE-Net) accurately classifies mammograms for breast cancer detection without manual region annotation. This AI tool aids radiologists by improving diagnostic accuracy and identifying suspicious areas.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Early breast cancer detection is crucial for effective treatment.
- Convolutional neural networks (CNNs) show promise in mammogram analysis, but challenges remain in classifying lesions of varying sizes and extracting subtle features.
- Current methods often rely on labor-intensive region of interest (ROI) annotation, hindering scalability.
Purpose of the Study:
- To develop a novel network for automated mammogram classification (cancer vs. no cancer) that overcomes limitations of existing computer-aided diagnosis systems.
- To improve the accuracy and efficiency of mammogram interpretation, assisting radiologists in early breast cancer diagnosis.
Main Methods:
- Proposed a novel Feature Selection and Enhancement Network (FSE-Net) that utilizes only mammogram images and image-level labels, eliminating the need for bounding boxes or masks.
- Implemented an adaptive feature selection module to fuse features from multi-scale receptive fields, capturing contextual information.
- Designed a feature enhancement module to leverage correlations between feature maps of different resolutions, improving representational capacity.
Main Results:
- On the CBIS-DDSM dataset, FSE-Net achieved an accuracy of 0.806 and an Area Under the Curve (AUC) of 0.866.
- On the INbreast dataset, FSE-Net demonstrated superior performance with an accuracy of 0.956 and an AUC of 0.974.
- Saliency map analysis indicated the network can roughly locate suspicious regions, aiding in final image prediction.
Conclusions:
- The proposed FSE-Net achieves satisfactory performance for mammogram classification, demonstrating its potential as an effective tool for breast cancer diagnosis.
- The network's ability to automatically classify mammograms and roughly identify suspicious regions can assist radiologists, potentially leading to more accurate and efficient diagnoses.
- FSE-Net addresses the challenges of lesion size variation and feature extraction without requiring laborious manual annotations.
