PaperlayerView exact revision on arXiv

Source-linked AI summary

Multivariate mixture model for myocardium segmentation combining multi-source images

Xiahai Zhuang

arXiv:1612.08820v1cs.CV

TL;DR

Indistinguishable myocardial boundaries and misaligned multi-source CMR images complicate segmentation. The paper introduces MvMM-based simultaneous segmentation with spatial transformations, reporting statistically significant improvements over conventional approaches and noting limitations in image choice and automation.

  • Problem

    Myocardial boundaries can be indistinguishable from adjacent regions because infarcted myocardium may resemble blood pools and healthy myocardium has overlapping intensity ranges.

  • Method

    The method uses a multivariate mixture model for simultaneous multi-source image segmentation, representing images in a common spatial domain and using multivariate image vectors.

  • Results

    The proposed MvMM significantly outperformed conventional approaches, including Atlas+bSSFP and GMM+bSSFP, with p<0.01; reported Dice scores were 0.866 for endocardium, 0.896 for epicardium, and 0.717 for myocardium.

  • Takeaways & Limitations

    MvMM provides a unified approach for simultaneous segmentation of multi-source images and improves segmentation performance relative to methods evaluated separately or with alternative combinations.

  • Takeaways & Limitations

    The work identifies T2 CMR as a potentially sub-optimal choice and notes that using bSSFP shape information is neither fully automatic nor integrated into a unified framework.

Abstract

from arXiv · show

This paper proposes a method for simultaneous segmentation of multi-source images, using the multivariate mixture model (MvMM) and maximum of log-likelihood (LL) framework. The segmentation is a procedure of texture classification, and the MvMM is used to model the joint intensity distribution of the images. Specifically, the method is applied to the myocardial segmentation combining the complementary texture information from multi-sequence (MS) cardiac magnetic resonance (CMR) images. Furthermore, there exist inter-image mis-registration and intra-image misalignment of slices in the MS CMR images. Hence, the MvMM is formulated with transformations, which are embedded into the LL framework and optimized simultaneously with the segmentation parameters. The proposed method is able to correct the inter- and intra-image misalignment by registering each slice of the MS CMR to a virtual common space, as well as to delineate the indistinguishable boundaries of myocardium consisting of pathologies. Results have shown statistically significant improvement in the segmentation performance of the proposed method with respect to the conventional approaches which can solely segment each image separately. The proposed method has also demonstrated better robustness in the incongruent data, where some images may not fully cover the region of interest and the full coverage can only be reconstructed combining the images from multiple sources.

1 INTRODUCTION

Myocardial segmentation is clinically important but difficult to automate because pathological and healthy tissues can have overlapping, heterogeneous, and complex intensities. The paper motivates simultaneous analysis of misaligned, differently covered multi-source CMR images to combine complementary information.

  • Multi-sequence CMR provides complementary information: LGE visualizes infarction, T2 images acute injury and ischemic regions, and bSSFP captures motion with clear boundaries.
  • Automated myocardial segmentation is sought because manual delineation is time-consuming, tedious, and subject to inter- and intra-observer variation.
  • Overlapping intensities can make myocardial boundaries indistinguishable from adjacent organs, regions, or blood pools in LGE CMR.Infarcted myocardium may resemble blood pools, while healthy myocardium may resemble liver or lung.
  • Pathologies create heterogeneous myocardial intensities and complex, patient-dependent enhancement patterns that invalidate simple distribution assumptions.Microvascular obstruction can appear as hypo-enhanced areas within hyper-enhanced infarcted regions.
  • Existing approaches often segment images separately or propagate one image’s result, making consistent and accurate multi-source segmentation challenging.Prior methods include registration, templates, shape constraints, and separate use of bSSFP-derived anatomical guidance.
  • The proposed unified framework simultaneously segments multi-source images using an MvMM and maximum log-likelihood framework while addressing inter-image and intra-image slice misalignment.The approach combines complementary information and formulates segmentation with transformations for multi-source CMR.

2 METHOD

The method jointly segments aligned multi-source images by modeling their combined intensity information with an MvMM, while optimizing registration transformations in a log-likelihood framework. It supports complementary texture information, corrects misalignment, and handles images with incomplete coverage.

  • Multivariate image modeling: The method classifies anatomical structures from combined multi-source medical images using a multivariate image variable.The approach is motivated by modalities that provide diverse morphological and pathological information.
  • Multivariate image modeling: Shared tissue labels can coexist with image-specific subtype and texture classifications, allowing complementary structures to be represented across modalities.For example, scars may be distinguishable in LGE but not in bSSFP or T2 images.
  • Multivariate image modeling: An MvMM models the joint intensity distribution of images acquired from the same subject in a shared common space.The common space is the region-of-interest coordinate system defined by combining the multi-source images.
  • Parameter estimation: The model estimates hidden labels and component information with an EM procedure that iteratively updates model parameters and segmentation variables.The generalized EM scheme guarantees an improvement of the likelihood at each iteration.
  • Registration: Affine slice transformations and atlas deformation are optimized alongside segmentation parameters to correct motion shifts, inter-image misregistration, and local atlas misregistration.The framework alternates parameter groups using coordinate ascent until a local optimum is found.
  • Hetero-coverage images: The MvMM formulation is generalized to hetero-coverage multivariate images when some image content is not acquired.This formulation addresses differing image resolutions and region-of-interest coverage.

3 EXPERIMENTS AND RESULTS

The experiments evaluate the proposed MvMM for multi-sequence CMR myocardial segmentation, compare it with conventional and alternative-registration methods, and test its behavior under different model dimensions and coverage conditions. The supplied passages describe the experimental design and datasets but do not report numerical outcomes.

  • Experimental design: Three experiments evaluate MvMM myocardial segmentation from multi-sequence CMR, including data, metrics, implementation, and variation analyses.The section studies inter-observer and inter-sequence variations.
  • Experimental design: The second experiment compares conventional segmentation approaches with MvMM using alternative registration schemes, focusing on challenging LGE CMR.LGE CMR is identified as the main research focus because its segmentation is challenging.
  • Experimental design: The third experiment studies MvMM performance with different multivariate-variable dimensions and hetero-coverage MVIs using simulated BrainWeb images.The study varies both the number of images and region-of-interest coverage.
  • Data description: The supplied experiment passages identify Table 1 as summarizing the parameters of the three CMR sequences.No quantitative results are stated in the provided passages.

3.1 Segmentation Combining Multi-Sequence CMR

The study combines LGE, T2, and bSSFP CMR sequences in an MvMM framework to segment myocardium using complementary intensity information while accommodating inconsistent coverage and alignment. Visual and quantitative assessments report accurate, robust segmentation, especially for challenging LGE images.

  • Data: The dataset contains LGE, T2, and bSSFP CMR sequences with substantially different slice coverage, resolutions, and thicknesses.LGE covers the ventricles with more than ten 5 mm slices, T2 has only three to seven 12–20 mm slices, and bSSFP fully covers the ventricles.
  • Evaluation: The evaluation compares automatic segmentations with gold standards formed by averaging three independent manual slice-by-slice delineations using Dice and ACD.Dice measures volume overlap, whereas lower ACD indicates better contour accuracy.
  • Atlas: The atlas is constructed from manually segmented healthy-subject bSSFP images that are nonrigidly registered to a reference.Atlas intensity and label-probability images are computed from transformed images and label information, with Gaussian-smoothed normalized label probabilities.
  • Method: The MvMM registers the atlas to whole-ventricle bSSFP and propagates atlas probabilities into a common space shared by the multi-sequence CMR images.The registration scheme uses affine, locally affine, and free-form deformation transformations, while modeling multiple tissue components across sequences.
  • Visual assessment: The proposed MvMM achieves good accuracy and robustness across the three sequences, particularly for the challenging LGE myocardial-segmentation task.Visual examples show accurate delineation of myocardial scars and boundaries that are indistinguishable from blood pools in LGE images.
  • Quantitative assessment: T2 segmentation scores are lower than bSSFP scores, while adding GMM-based assistance increases the myocardium Dice score mainly through improved endocardial delineation.The epicardial result worsens in this comparison.

3.2 LGE CMR Segmentation Using Different Schemes

This section compares separate and combined segmentation schemes for LGE CMR, including registration corrections for atlas alignment and shifted slices. The proposed MvMM benefits from information in the other sequences and from correcting inter-slice motion.

  • Separate segmentation: Four conventional schemes separately segment LGE CMR using Atlas, GMM, Atlas+bSSFP, or GMM+bSSFP initialization.The latter two use the subject’s bSSFP image to assist atlas-to-target registration or initialize GMM.
  • Combined segmentation: The combined experiments evaluate Mvmm⊖ without registration correction, Mvmm⊖+SC with shift correction, and MvMM with both FFD and SC correction.The proposed MvMM combines free-form deformation and shift correction.
  • Quantitative comparison: The four separate segmentation schemes obtain significantly worse myocardium Dice scores than the proposed combined MvMM (p < 0.01).Both Atlas+bSSFP and GMM+bSSFP also outperform their counterparts without bSSFP assistance (p < 0.01).
  • Visual assessment: The proposed MvMM produces better LGE myocardial segmentation than GMM by using guidance and constraints from the other two sequences.This is especially relevant in scar regions where the LGE boundary is poorly visible or absent.

3.3 Study MvMM Using Simulated Brain MR

Using simulated multi-sequence brain MR images, the study tests how the number of images and incomplete region-of-interest coverage affect MvMM segmentation. More image dimensions improve Dice performance, and combined schemes remain robust under substantial truncation.

  • Experimental design: The brain experiment uses T1-weighted, T2-weighted, and proton-density MR images with simulated inter-sequence mis-registration.Fifty cases are generated by deforming original images so they are slightly misaligned with probabilistic atlases.
  • Experimental design: The comparison includes UvMM1 using one image, MvMM2 using two images, and MvMM3 using three images.The experiment evaluates MvMM performance as the multivariate image dimension increases.
  • Image number: Dice performance improves as more images form the multivariate image, with a larger gain from univariate to bivariate than from bivariate to trivariate.The results indicate performance tends to converge after a certain number of images are combined.
  • Incomplete coverage: MvMM2 and MvMM3 retain good performance when images are randomly truncated by more than 40 mm of slices.This supports greater robustness to incomplete region-of-interest coverage than single-image segmentation.
  • Segmentation performance: MvMM2 and MvMM3 achieve much higher white- and gray-matter Dice scores than UvMM1.The Dice scores are reported for the segmentation of both tissue classes.
  • Incomplete coverage: The combined segmentation improves robustness when some images do not fully cover the region of interest.Fuller coverage can be obtained by combining multi-source images.

4 CONCLUSION AND DISCUSSION

The paper presents MvMM for simultaneous segmentation of multi-source images, aligning MS CMR slices through transformations while combining complementary image information. It reports improved myocardial segmentation, robustness to incomplete ROI coverage, and broader applicability to multi-source segmentation and classification.

  • Method: MvMM simultaneously segments multi-source images by modeling their joint information and registering MS CMR slices to a common space.The MvMM and transformation variables are optimized within the log-likelihood framework using separate parameter-update procedures.
  • Myocardial segmentation results: 0.866, 0.896, and 0.717 were the reported LGE CMR Dice scores for endocardium, epicardium, and myocardium, respectively.The corresponding ACD values for endocardium and epicardium were 2.17 mm and 2.16 mm; these were compared with inter-observer values of 0.876, 0.903, and 0.757 for the three Dice scores.
  • Myocardial segmentation results: The proposed method significantly outperformed separate atlas-based and GMM segmentations on LGE CMR, including versions initialized with bSSFP registration (p<0.01).Using bSSFP improved separate-segmentation Dice scores, but those results remained significantly worse than MvMM.
  • Registration: Registration generally improved MvMM segmentation, with motion-shift correction producing a more evident improvement than atlas-to-target registration.The comparison included MvMM without registration, with nonrigid atlas-to-target registration, and with affine registration.
  • Multi-source robustness: Combined MvMM segmentation was more robust than separate UvMM1 segmentation when some images had incomplete ROI coverage.Higher-dimensional MvMM performance became comparable after combining two-source images, indicating convergence in that experiment.
  • Limitations and broader applicability: The main scope limitations are sparse T2 CMR coverage, possible nonrigid inter-sequence misalignment, and the need to extend deformation modeling in future work.The authors also describe MS CMR myocardial segmentation as an example application and suggest broader use across multi-modality, multi-source imaging tasks.
Loading 1612.08820v1…