Computerized Lesion Segmentation on DCE-MRI using Active Contours and Spectral Embedding

Introduction: Accurate lesion segmentation is an important component of determining quantitative features for lesions on MRI. In this study, we develop an automated lesion segmentation method for delineating lesions on dynamic contrast enhanced (DCE)-MRI. We present a new active contour model which uses spectral embedding (SE), a process that partitions images in order to maximize intercluster similarity and minimize intracluster similarity while preserving object adjacencies [1]. SE has been used previously for classification of high dimensional data [2], but this is the first known application of SE in the context of segmentation. SE transforms the images into a data representation that accentuates the gradients at the lesion borders which allows the active contour to accurately identify the lesion boundaries. This will allow better characterization of lesion morphology, which facilitates discrimination between lesions with subtle shape differences. We demonstrate on a cohort of 50 breast lesions imaged on DCE-MRI that automated segmentation using our spectral embedding based active contour (SEAC) model is more similar to the manual lesion segmentation performed by a radiologist who is a breast imaging specialist than the popular fuzzy c-means (FCM) method [3] combined with an active contour model. Methods: DCE-MRIs of 50 breast lesions from 50 patients were collected under IRB approval. Sagittal T1‐weighted, spoiled gradient echo sequences with fat suppression consisting of one series before contrast injection of Gd‐DTPA (precontrast) and 3‐8 series after contrast injection (postcontrast) were acquired at either 1.5 Tesla or 3 Tesla (Siemens Magnetom or Trio, respectively). Single slice dimensions were 384x384, 512x512, or 896x896 pixels with a slice thickness of 3 mm. Temporal resolution between postcontrast acquisitions was a minimum of 90 seconds. An attending radiologist selected the 2D slice that was most representative of the lesion, and the radiologist manually delineated the boundary of the lesion. SE is performed along the time dimension to reduce the dimensionality of the time domain from 4-9 (i.e., total number of time points) to 3. Each resulting dimension is then represented as one color channel of a hue/saturation/value (HSV) color space. A color active contour method is then evolved on the embedding color map. For the comparison FCM method [3], voxels in each image were clustered into 3 data classes. The 3 channel color map resulting from these 3 data classes was then used in conjunction with the same color active contour method used with SEAC.