Visual exploratory analysis of DCE-MRI data in breast cancer by dimensional data reduction: A comparative study

One trend in modern medical imaging is the growing signal dimension in new multi-modal or multivariate imaging approaches. To analyze such high dimensional data, new approaches need to be proposed and evaluated. The scope of this study is to investigate the potential of three different algorithms for dimensional data reduction for the visual exploration of biomedical signals arising from dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) applied to breast cancer detection. The algorithms employed are the established Principal Component Analysis (PCA) and Self-Organizing Maps (SOM), and the recently proposed Locally Linear Embedding (LLE). The experimental dataset comprises the time-series associated with the voxels of six benign and six malignant breast tumors. In order to visually explore the dataset, the multi-dimensional signal space of all the time-series is projected into a two-dimensional space by PCA, SOM and LLE, respectively. We show how the visualization of the respective projected spaces with customized colors can allow the user to discover hidden regularities in the data, in particular with regard to the differentiation between benign and malignant lesions. The performances of the three algorithms are quantitatively compared, while discussing their advantages and drawbacks.