|
Abstract:
|
Conventionally, radiologists exam intratumoral heterogeneity by the visual presentation of multiple images from Dynamic contrast-enhanced MRI (DCE-MRI). We propose a new clustering algorithm that summaries tumor vacularity while incorporates multivariate measurements and spatial relationship. We used a multivariate mixture-model-based clustering method. Taking the advantage of Bayesian paradigm, the class labels are modeled via a latent process with its prior governed by a CAR model. Nearby pixels are more likely to be grouped into the same class. A grid image was divided into 9 sub-regions. Each sub-region was pre-assigned to one of the three classes. We compared our proposed Bayesian cluster method to the fuzzy c-means (non-spatial method) results.Our method had higher Cohen's Kappa and lower MSE. The latent process (v) also provided a spatially correlated heat map of the pixel enhancements. Applied to 24 kidney tumors, our method showed smoother partitioning with less noise than FCM. Our study provides an improved cluster method that takes advantage of spatial information.
|
