Multi-atlas propagation via a manifold graph on a database of both labeled and unlabeled images

We present a framework for multi-atlas based segmentation in situations where we have a small number of segmented atlas images, but a large database of unlabeled images is also available. The novelty lies in the application of graph-based registration on a manifold to the problem of multi-atlas registration. The approach is to place all the images in a learned manifold space and construct a graph connecting near neighbors. Atlases are selected for any new image to be segmented based on the shortest path length along the manifold graph. A multi-scale non-rigid registration takes place via each of the nodes on the graph. The expectation is that by registering via similar images, the likelihood of misregistrations is reduced. Having registered multiple atlases via the graph, patch-based voxel weighted voting takes place to provide the nal segmentation. We apply this approach to a set of T2 MRI images of the prostate, which is a notoriously di cult segmentation task. On a set of 25 atlas images and 85 images overall, we see that registration via the manifold graph improves the Dice coe cient from 0.82±0.05 to 0.86±0.03 and the average symmetrical boundary distance from 2.89±0.62mm to 2.47± 0.51mm. This is a modest but potentially useful improvement in a di cult set of images. It is expected that our approach will provide similar improvement to any multi-atlas segmentation task where a large number of unsegmented images are available. 1. DESCRIPTION OF PURPOSE Prostate segmentation is important in providing models for surgical guidance and for diagnosis and staging of disease. Quantitative analysis of the e ectiveness of such techniques requires a large number of accurate segmentations. However, such accurate anatomical annotations are di cult to obtain. The only truly reliable segmentation method is manual delineation by well-trained radiologists. This is a very time-consuming process and is not feasible in normal clinical practice. As a result, automatic segmentation is a growing research eld. Among these methods, the most robust and accurate segmentations are based on empirical information learned from manually labeled training sets. Multi-atlas segmentation has proved successful for images such as brain MRI and abdominal CT. It can also be applied to T2 MRI of the prostate, but there is large variation in both the shape and intensity of the prostate in such scans. An accurate, generic, robust and automated segmentation remains a signi cant research challenge. In order to tackle this problem, several approaches have been explored for multi-atlas segmentation. Atlases in the database which are the most similar to the query image should be used for more accurate registration. We previously proposed appearance-speci c atlases and a locally weighted atlas segmentation to achieve accurate delineation of the local boundary. In most circumstances, it is too time consuming to obtain enough labeled atlases. However, a large number of unsegmented images may be more readily available. To utilize unlabeled images, Wolz et al. proposed a novel manifold learning embedding method to place all the images into the manifold space, and an initial set of atlases is propagated to all images through a succession of multi-atlas segmentation steps. However, the propagation error can be signi cantly accumulated if the labeled atlases are insu cient. Rubinstein et al. incorporated dense image correspondence of all labeled and unlabeled images to infer annotations automatically in a large database. Cao et al. suggested manifold embedding for atlas segmentation of prostate MRI images. Their method is only demonstrated on 2D images and is used for atlas selection. There is no incorporation of unsegmented images. Email for correspondence: [email protected], [email protected] . Medical Imaging 2014: Computer-Aided Diagnosis, edited by Stephen Aylward, Lubomir M. Hadjiiski, Proc. of SPIE Vol. 9035, 90350A · © 2014 SPIE · CCC code: 1605-7422/14/$18 · doi: 10.1117/12.2044027 Proc. of SPIE Vol. 9035 90350A-1 Downloaded From: http://spiedigitallibrary.org/ on 04/08/2014 Terms of Use: http://spiedl.org/terms Figure 1. The example of registration via intermediate images with small variation Graph based image registration has been investigated to generate more accurate transformations between image pairs with considerable anatomical variation. Geodesic estimation and geodesic registration has also been proposed to solve the large deformation registration problem. Cardoso et al. propose a method for propagation of information across brain images using a morphological and intensity based local manifold space. However, these methods have not yet been extended to multi-atlas segmentation. In non-rigid image registration, it is challenging to achieve an accurate transformation between images with large variation. Using B-spline based non-rigid registration, there may be thousands of control points which mean a large number of degrees of freedom of the cost function. If two images are far way, it is easy to fall into local minima during optimization. However, the registration between images with small deformation performs well. Fig. 1 shows that registering a oating image via intermediate images with small variation to a target image performs much better than directly registering from the oating image to the target image. This suggests a possible approach to deal with our prostate registration problem. We propose a practical segmentation tool for T2 prostate MRI images. The method relies on a large number of unsegmented images being available, but only a modest number of segmented atlases. The manifold embedding requires only a ne registration and simple coarse region of interest estimation. The aim is to provide accurate segmentation of unseen images in a practical timescale.