Multi-atlas Segmentation with Robust Label Transfer and Label Fusion

Abstract

Multi-atlas segmentation has been widely applied in medical image analysis. This technique relies on image registration to transfer segmentation labels from pre-labeled atlases to a novel target image and applies label fusion to reduce errors produced by registration-based label transfer. To improve the performance of registration-based label transfer against registration errors, our first contribution is to propose a label transfer scheme that generates multiple warped versions of each atlas to one target image through registration paths obtained by composing inter-atlas registrations and atlas-target registrations. The problem of decreasing quality of warped atlases caused by accumulative errors in composing multiple registrations is properly addressed by an atlas selection method that is guided by atlas segmentations. To improve the performance of label fusion against registration errors, our second contribution is to integrate the probabilistic correspondence model employed by the non-local mean approach with the joint label fusion technique, both of which have shown excellent performance for label fusion. Experiments on mitral-valve segmentation in 3D transesophageal echocardiography (TEE) show the effectiveness of the proposed techniques.

Multi-atlas segmentation of biomedical images: A survey

“…As explained in Section 2.1 above, generating more candidate segmentations can improve the subsequent fusion, but also might worsen it by introducing poor candidates generated by suboptimal registrations. A similar strategy, proposed by Wang et al (2013a), employs pre-computed registrations between pairs of atlases to generate a multitude of propagated labels by concatenating the pairwise registration results. In a parallel effort, Datteri et al (2014) relied on pre-registered atlases to estimate registration accuracy for the novel image based on registration circuits.…”

“…Finally, there are many other applications that have benefited from MAS within human medical imaging, including: segmentation of pelvic bones in MRI (Weisenfeld and Warfield, 2011b; Akhondi-Asl et al, 2014); lungs in CT scans (van Rikxoort et al, 2009) and chest X-rays (Candemir et al, 2014); heart and its ventricles in CT (van Rikxoort et al, 2010; Dey et al, 2010), MRI (Zhuang et al, 2010; Zuluaga et al, 2014), MR angiography (Wachinger and Golland, 2012), ultrasound (Wang et al, 2014a), and CT angiography (Kirişli et al, 2010; Yang et al, 2014a); breast tissues and lesions in X-ray mammography (Iglesias and Karssemeijer, 2009) and MRI (Gubern-Mérida et al, 2012; Lee et al, 2013); cartilage and bone in knee MRI (Tamez-Pena et al, 2012; Lee et al, 2014b; Shan et al, 2014); the vertebrae in spinal MRI (Asman et al, 2014); scar tissue in intravascular coronary optical coherence tomography (OCT) (Tung et al, 2013); the mitral valve in transesophageal echocardiography (Wang et al, 2013a; Pouch et al, 2014); skeletal muscle in whole-body MRI (Karlsson et al, 2014); kidneys in CT images (Yang et al, 2014b); and bone in dental cone-beam CT images (Wang et al, 2014c). …”

Multi-atlas segmentation with joint label fusion and corrective learning—an open source implementation

“…Note that, similar to the non-local mean patch based label fusion approach Coupe et al (2011), employing all patches within the searching neighborhood for estimating the pairwise atlas dependencies produces more accurate estimation Wang et al (2013a). However, this approach has much higher computational complexity.…”

“…Under the assumption that segmentation errors produced by different atlases are not identical, it is often feasible to derive more accurate solutions by label fusion. Since the example-based knowledge representation and registration-based knowledge transfer scheme can be effectively applied in many biomedical imaging problems, label fusion based multi-atlas segmentation has produced impressive automatic segmentation performance for many applications (Rohlfing et al, 2004; Isgum et al, 2009; Collins and Pruessner, 2010; Asman and Landman, 2012; Wang et al, 2013a). For some most studied brain image segmentation problems, such as hippocampus segmentation (Wang et al, 2011) and hippocampal subfield segmentation (Yushkevich et al, 2010), automatic segmentation performance produced by multi-atlas label fusion has reached the level of inter-rater reliability.…”

MUSE: MUlti-atlas region Segmentation utilizing Ensembles of registration algorithms and parameters, and locally optimal atlas selection

“…In this way, we obtain a large ensemble of tentative label maps that are generated by applying a multitude of transformations on multiple atlases, and we use the ensemble for deriving final labels for each voxel. The general concept of generating a larger ensemble of label maps was explored in a few recent papers: in Wang et al (2013) multiple warps from the same atlas were generated by composing inter-atlas registrations and atlas-target registrations; in Pipitone et al (2014) segmentations from a small number of atlases were propagated to a subset of target images and the new atlases were used for segmenting all target images. However, these methods used a different approach than ours, by following an “atlas propagation” strategy.…”