Feature Extraction and Classification of Dynamic Contrast-Enhanced T2-weighted Breast Image Data


UNLABELLED The relatively low specificity of dynamic contrast-enhanced T1-weighted magnetic resonance imaging (MR) imaging of breast cancer has lead several groups to investigate different approaches to data acquisition, one of them being the use of rapid T2*-weighted imaging. Analyses of such data are difficult due to susceptibility artifacts and breathing motion. MATERIALS AND METHODS One-hundred-twenty-seven patients with breast tumors underwent MR examination with rapid, single-slice T2*-weighted imaging of the tumor. Different methods for classifying the image data set using leave-one-out cross validation were tested. Furthermore, a semi-automatic region of interest (ROI) definition tool was presented and compared with manual ROI definitions from a previous study. Finally, pixel-by-pixel analysis was done and compared with ROI analysis. The analyses were done with and without noise reduction. RESULTS The minimum enhancement parameter was the most robust and accurate of the parameters tested. The semi-automatic ROI definition method was fast and produced similar results as the manually defined ROIs. Noise reduction improved both sensitivity and specificity, but the improvement was not statistically significant. The pixel-based analysis methods used in the present study did not improve classification results. CONCLUSIONS Analysis of T2*-weighted breast images can be done in a rapid and robust manner by using semi-automatic ROI definition tools in combination with noise reduction. Minimum enhancement gives an indication of malignancy in T2*-weighted imaging.

DOI: 10.1109/42.974924

Cite this paper

@article{Torheim2001FeatureEA, title={Feature Extraction and Classification of Dynamic Contrast-Enhanced T2-weighted Breast Image Data}, author={Geir Torheim and Fred Godtliebsen and David Axelson and Kjell Arne Kvistad and Olav Haraldseth and Peter A. Rinck}, journal={IEEE transactions on medical imaging}, year={2001}, volume={20 12}, pages={1293-301} }