Lesion Segmentation in Dynamic Contrast Enhanced MRI of Breast

Dynamic contrast enhanced magnetic resonance imaging (DCE-MRI) is a sensitive tool used for the detection of breast cancer. Automated segmentation of breast lesions in DCE-MR images is challenging due to the inherent low signal-to-noise ratios and high inter-patient variability. A lesion segmentation method based on supervised classification is proposed in this study. In this method, a DCE-MR image is modeled as a connected graph with local Markov properties where each voxel of the image is regarded as a node. Two kinds of edge potentials of the graph are proposed to encourage the smoothness and continuity of the segmented regions. In the supervised classification based lesion segmentation of the DCE-MRI, one main difficulty is that the levels and ranges of intensities and enhancement features can vary significantly among patients. For instance, the normal parenchymal tissues of a patient may present a similar enhancement pattern or level as the lesion tissues in another patient. We propose a robust normalization method on the intensity and kinetic features such that the feature values in different MR images are similar in terms of scales and ranges. The segmentation schemes with the two proposed edge potentials show significantly higher lesion overlap rates with the ground truth of 51 +/- 26% and 48 +/- 25% on 30 lesions respectively, compared to the fuzzy c-means of 6 +/- 9% (baseline) and a recently proposed multi-channel Markov random field of 36 +/- 23%. Our methods have consistently outperformed the existing methods on cases with mild, moderate, marked and mixed background enhancement.