Related Experiment Video
Updated: Mar 15, 2026
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.6K
Digital mammographic tumor classification using transfer learning from deep convolutional neural networks.
Benjamin Q Huynh1, Hui Li1, Maryellen L Giger1
1University of Chicago , Department of Radiology, 5841 South Maryland Avenue, Chicago, Illinois 60637, United States.
Journal of Medical Imaging (Bellingham, Wash.)
|September 10, 2016
Summary
Transfer learning with Convolutional Neural Networks (CNNs) shows promise for breast cancer diagnosis. This approach effectively extracts tumor features from mammograms, comparable to traditional methods, and improves diagnostic accuracy when combined with existing techniques.
Area of Science:
- Radiology and Medical Imaging
- Machine Learning in Healthcare
- Computational Pathology
Background:
- Convolutional Neural Networks (CNNs) offer potential for computer-aided diagnosis (CADx) by directly learning image features.
- Training CNNs from scratch for medical images is challenging due to limited datasets and presentation variability.
- Transfer learning leverages models pretrained on nonmedical tasks to overcome data scarcity in medical imaging.
Purpose of the Study:
- To evaluate the efficacy of transfer learning using CNNs for extracting tumor features from mammographic images.
- To compare the performance of machine learning classifiers based on CNN-extracted features versus traditional computer-extracted features for breast lesion classification.
- To assess the diagnostic performance of ensemble classifiers combining both feature types.
Main Methods:
- Utilized a dataset of 219 breast lesions from 607 full-field digital mammographic images.
- Employed transfer learning with CNNs pretrained on nonmedical tasks to extract image features.
- Compared Support Vector Machine (SVM) classifiers using CNN-extracted features and previously defined computer-extracted features.
- Conducted five-fold cross-validation using the area under the receiver operating characteristic (ROC) curve as the performance metric.
Main Results:
- Classifiers utilizing CNN-extracted features via transfer learning demonstrated performance comparable to those using analytically extracted features (AUC [Formula: see text]).
- Ensemble classifiers integrating both CNN-extracted and analytically extracted features significantly outperformed classifiers using either feature type alone ([Formula: see text] vs. 0.81, [Formula: see text]).
Conclusions:
- Transfer learning offers a viable method to enhance current CADx systems for breast lesion classification without requiring extensive medical datasets.
- This approach facilitates the development of standalone classifiers and integrates effectively with existing radiomic methods.
- The findings support the application of machine learning in radiomics and precision medicine for improved diagnostic capabilities.