Related Experiment Video
Updated: Jun 16, 2026
15:48
Tracking the Mammary Architectural Features and Detecting Breast Cancer with Magnetic Resonance Diffusion Tensor Imaging
Published on: December 15, 2014
Exploring nonlinear feature space dimension reduction and data representation in breast Cadx with Laplacian eigenmaps
Andrew R Jamieson1, Maryellen L Giger, Karen Drukker
1Department of Radiology, University of Chicago, Chicago, Illinois 60637, USA. andrewj@uchicago.edu
Medical Physics
|February 24, 2010
Summary
New nonlinear dimension reduction techniques show promise for breast cancer diagnosis, matching or exceeding current methods. These methods offer complementary data visualization for improved interpretation in computer-aided diagnosis.
Area of Science:
- Medical Imaging
- Machine Learning
- Computer-Aided Diagnosis
Background:
- Computer-extracted features from medical images are crucial for breast cancer diagnosis.
- Traditional methods for analyzing these high-dimensional feature spaces have limitations.
Purpose of the Study:
- To evaluate novel unsupervised nonlinear dimension reduction (DR) techniques for breast lesion feature spaces.
- To compare these methods against existing computer-aided diagnosis (CADx) algorithms across Ultrasound, MRI, and mammography data.
Main Methods:
- Applied Laplacian eigenmaps and t-distributed stochastic neighbor embedding (t-SNE) for nonlinear DR.
- Evaluated classification performance using Markov chain Monte Carlo based Bayesian artificial neural network (MCMC-BANN) and linear discriminant analysis.
- Compared results with principal component analysis (PCA), automatic relevance determination (ARD), and linear stepwise (LSW) feature selection.
Main Results:
- Nonlinear DR techniques, particularly 4D t-SNE, achieved high performance (AUC 0.90) in Ultrasound data using MCMC-BANN.
- These methods demonstrated comparable or superior classification performance to established feature selection techniques.
- The techniques provided sparse, interpretable lower-dimensional representations of the feature space.
Conclusions:
- Novel nonlinear DR methods show potential to enhance breast lesion CADx performance.
- These DR techniques can complement existing feature selection methods.
- The ability to generate sparse, visually interpretable data representations is a key benefit for data analysis.
