GPSSI: Gaussian Process for Sampling Segmentations of Images
Medical image segmentation is often a prerequisite for clinical applications. As an ill-posed problem, it leads to uncertain estimations of the region of interest which may have a significant impact on downstream applications, such as therapy planning. To
- PDF / 1,612,338 Bytes
- 9 Pages / 439.363 x 666.131 pts Page_size
- 9 Downloads / 184 Views
2
Asclepios Project, INRIA Sophia Antipolis, France Department of Radiation Oncology, Massachusetts General Hospital and Harvard Medical School, Boston, MA, USA
Abstract. Medical image segmentation is often a prerequisite for clinical applications. As an ill-posed problem, it leads to uncertain estimations of the region of interest which may have a significant impact on downstream applications, such as therapy planning. To quantify the uncertainty related to image segmentations, a classical approach is to measure the effect of using various plausible segmentations. In this paper, a method for producing such image segmentation samples from a single expert segmentation is introduced. A probability distribution of image segmentation boundaries is defined as a Gaussian process, which leads to segmentations that are spatially coherent and consistent with the presence of salient borders in the image. The proposed approach outperforms previous generative segmentation approaches, and segmentation samples can be generated efficiently. The sample variability is governed by a parameter which is correlated with a simple DICE score. We show how this approach can have multiple useful applications in the field of uncertainty quantification, and an illustration is provided in radiotherapy planning.
1
Introduction
Medical image analysis, and in particular medical image segmentation, is a key technology for many medical applications, ranging from computer aided diagnosis to therapy planning and guidance. Medical image segmentation is probably the task most often required in those applications. Due to its ill-posed nature, the quantification of segmentation accuracy and uncertainty is crucial to assess the overall performance of other applications. For instance, in radiotherapy planning it is important to estimate the impact of uncertainty in the delineation of the gross tumor volume and the organs at risk on the dose delivered to the patient. A straightforward way to assess this impact is to perform Image Segmentation Sampling (ISS), which consists of gathering several plausible segmentations of the same structure, and estimate the variability of the output variables due to the variability of the segmentations. For computer generated segmentations, ISS could simply be obtained by varying the parameters or initial values of the algorithm producing the segmentations. However, in many cases, parameters of the algorithms cannot be modified, and segmentations are partially edited by a user. For manual or semi-manual segmentations, it is possible to estimate the c Springer International Publishing Switzerland 2015 N. Navab et al. (Eds.): MICCAI 2015, Part III, LNCS 9351, pp. 38–46, 2015. DOI: 10.1007/978-3-319-24574-4_5
GPSSI: Gaussian Process for Sampling Segmentations of Images
39
inter-expert variability on a few cases but it usually cannot be applied on large databases due to the amount of resources required. This is why it is important to automate the generation of “plausible segmentations” that are “similar to” a given segmentation
Data Loading...
