Spectral Aggregation Based on Iterative Graph Cut for Sonographic Breast Image Segmentation

In this paper, an image segmentation framework is proposed by unifying the techniques of spectral clustering and graph-cutting to address the difficult problem of breast lesion demarcation in sonography. In order to alleviate the effect of speckle noise and posterior acoustic shadows, the ROI of a sonogram is mapped to a specific eigen-space as an eigenmap by a constrained spectral clustering scheme. The eigen-mapping is boosted with the incorporation of partial grouping setting and then provide a useful preliminary aggregation based on intensity affinity. Following that, an iterative graph cut framework is carried out to identify the object of interest in the projected eigenmap. The proposed segmentation algorithm is evaluated with four sets of manual delineations on 110 breast ultrasound images. The experiment results corroborates that the boundaries derived by the proposed algorithm are comparable to manual delineations and hence can potentially provide reliable morphological information of a breast lesion.