Robust unsupervised segmentation of infarct lesion from diffusion tensor MR images using multiscale statistical classification and partial volume voxel reclassification

Li, Wu; Tian, Jie; Li, Enzhong; Dai, Jianping

doi:10.1016/j.neuroimage.2004.08.009

Cited by 34 publications

(29 citation statements)

References 27 publications

Supporting

Mentioning

Contrasting

Order By: Relevance

“…As it is generally accepted that an SI greater than 0.7 indicates excellent agreement between detected lesions and true lesions [31], [42], [44], it is suggested that lesions with 40-80% intensity reduction were all well detected by the proposed method. This implied that our method was capable of identifying the major part of real lesions, with comparable accuracy to existing lesion detection methods using multispectral MRI scans [22]- [24], [43]. However, parts of lesion borders might be undetected if only minor lesions were included.…”

Section: Discussionmentioning

confidence: 68%

See 1 more Smart Citation

Detection of Infarct Lesions From Single MRI Modality Using Inconsistency Between Voxel Intensity and Spatial Location—A 3-D Automatic Approach

Shen

Szameitat

Sterr

2008

IEEE Trans. Inform. Technol. Biomed.

View full text Add to dashboard Cite

Abstract-Detection of infarct lesions using traditional segmentation methods is always problematic due to intensity similarity between lesions and normal tissues, so that multispectral MRI modalities were often employed for this purpose. However, the high costs of MRI scan and the severity of patient conditions restrict the collection of multiple images. Therefore, in this paper, a new 3-D automatic lesion detection approach was proposed, which required only a single type of anatomical MRI scan. It was developed on a theory that, when lesions were present, the voxel-intensitybased segmentation and the spatial-location-based tissue distribution should be inconsistent in the regions of lesions. The degree of this inconsistency was calculated, which indicated the likelihood of tissue abnormality. Lesions were identified when the inconsistency exceeded a defined threshold. In this approach, the intensity-based segmentation was implemented by the conventional fuzzy c-mean (FCM) algorithm, while the spatial location of tissues was provided by prior tissue probability maps. The use of simulated MRI lesions allowed us to quantitatively evaluate the performance of the proposed method, as the size and location of lesions were prespecified. The results showed that our method effectively detected lesions with 40-80% signal reduction compared to normal tissues (similarity index >0.7). The capability of the proposed method in practice was also demonstrated on real infarct lesions from 15 stroke patients, where the lesions detected were in broad agreement with true lesions. Furthermore, a comparison to a statistical segmentation approach presented in the literature suggested that our 3-D lesion detection approach was more reliable. Future work will focus on adapting the current method to multiple sclerosis lesion detection.Index Terms-Fuzzy c-mean (FCM), MRI, lesion detection, tissue probability map.

show abstract

Section: Discussionmentioning

confidence: 68%

“…Sensitivity measures a true positive detection rate, and specificity measures a true negative detection rate. Similarity index (SI) is derived from a reliability measure known as the kappa statistic, which is sensitive to both size and location of segmentation [23], [24], [31], [42], [43].…”

Section: B Performance Evaluation Parametersmentioning

confidence: 99%

Detection of Infarct Lesions From Single MRI Modality Using Inconsistency Between Voxel Intensity and Spatial Location—A 3-D Automatic Approach

Shen

Szameitat

Sterr

2008

IEEE Trans. Inform. Technol. Biomed.

View full text Add to dashboard Cite

show abstract

“…However, this approach is still based on a scalar measure, which neglects the orientation of the diffusion tensor. Li et al (11) suggested a multiscale statistical classification and partial volume voxel reclassification method, in which segmentation is performed in multiple stages on a stack of images at different levels of inner spatial scale.…”

Section: Introductionmentioning

confidence: 99%

A multivariate hypothesis testing framework for tissue clustering and classification of DTI data

et al. 2009

View full text Add to dashboard Cite

The primary aim of this work is to propose and investigate the effectiveness of a novel unsupervised tissue clustering and classification algorithm for diffusion tensor MRI (DTI) data. The proposed algorithm utilizes information about the degree of homogeneity of the distribution of diffusion tensors within voxels. We adapt frameworks proposed by Hext and Snedecor, where the null hypothesis of diffusion tensors belonging to the same distribution is assessed by an F-test. Tissue type is classified according to one of the four possible diffusion models, the assignment of which is determined by a parsimonious model selection framework based on Schwarz Criterion. Both numerical phantoms and diffusion-weighted imaging (DWI) data obtained from excised rat and pig spinal cords are used to test and validate these tissue clustering and classification approaches. The unsupervised clustering method effectively identifies distinct regions of interest (ROIs) in phantoms and real experimental DTI data.

show abstract

“…There are some semi-automatic and automatic segmentation algorithms [9][10][11][12]. Martel et al [9] report that their method works accurately (−2.45% and −1% error in infarct volume at high and low contrast, respectively, against manual segmentation errors of −2.24% and −4.7%) when the contrast between the stroke region and normal tissue is very good and the stroke region is large.…”

Section: Introductionmentioning

confidence: 99%

“…Martel et al [9] report that their method works accurately (−2.45% and −1% error in infarct volume at high and low contrast, respectively, against manual segmentation errors of −2.24% and −4.7%) when the contrast between the stroke region and normal tissue is very good and the stroke region is large. The method developed by Li et al [10,11] is based on a multi-stage integrated approach, which employs image processing, and atlas-based registration techniques. The method has been tested on 20 patients and the value of similarity index for three patients (about 6 slices) was from 0.931 to 0.978.…”

Section: Introductionmentioning

confidence: 99%

Automatic processing of diffusion-weighted ischemic stroke images based on divergence measures: slice and hemisphere identification, and stroke region segmentation

Prakash

Gupta

et al. 2008

Int J CARS

View full text Add to dashboard Cite

Objective Fast, accurate and automatic segmentation of acute ischemic stroke lesions is important for clinical trials and has potential for efficient stroke management. To identify stroke slices, stroke hemisphere, and segment stroke regions in diffusion-weighted magnetic resonance imaging (DWI), divergence based algorithms are proposed. Materials and methodsThe study used 57 DWI volumes with inter 0.94-2.42 mm and intra 5-14 mm plane resolutions. We used ratio of intensity probability density functions (pdf) as a divergence measure. For slice identification, this measure is the ratio of pdfs of slice and the volume; for hemisphere and infarct segmentation, it is the ratio of the difference of the pdfs of the left and right hemispheres to the sum of their pdfs. The median and cross over points are thresholds for slice and region segmentation while hemisphere identification is threshold free. Descriptive statistics were determined and ROC analysis was performed. Results The median sensitivity, specificity, and Dice statistical index for segmentation are 86.34%, 99.83%, 0.72, respectively. For slice and hemisphere identification sensitivity and specificity are (90.05%; 68.78%) and (94.74%; 94.74%), respectively. The algorithm implemented in VC++ takes 3-5 s per volume. Conclusion This automatic, accurate and fast method is potentially useful in clinical setting and clinical trials to localize and quantify the stroke regions, eliminate inter-and

show abstract

Robust unsupervised segmentation of infarct lesion from diffusion tensor MR images using multiscale statistical classification and partial volume voxel reclassification

Cited by 34 publications

References 27 publications

Detection of Infarct Lesions From Single MRI Modality Using Inconsistency Between Voxel Intensity and Spatial Location—A 3-D Automatic Approach

Detection of Infarct Lesions From Single MRI Modality Using Inconsistency Between Voxel Intensity and Spatial Location—A 3-D Automatic Approach

A multivariate hypothesis testing framework for tissue clustering and classification of DTI data

Automatic processing of diffusion-weighted ischemic stroke images based on divergence measures: slice and hemisphere identification, and stroke region segmentation

Contact Info

Product

Resources

About