Elsevier

Academic Radiology

Volume 13, Issue 12, December 2006, Pages 1474-1484
Academic Radiology

Original investigation
Identification, Segmentation, and Image Property Study of Acute Infarcts in Diffusion-Weighted Images by Using a Probabilistic Neural Network and Adaptive Gaussian Mixture Model

https://doi.org/10.1016/j.acra.2006.09.045Get rights and content

Rationale and Objectives

Accurate identification of infarcted regions of the brain is critical in management of stroke patients. An efficient and fast method for identification and segmentation of infarcts in the diffusion-weighted images (DWI) is proposed.

Materials and Methods

Thirteen stroke patients were studied. DWI scans were acquired with a slice thickness of 5 mm. We have used a probabilistic neural network for selecting infarct slices and an adaptive (two-level) Gaussian mixture model for segmentation of the infarcts. Statistical analysis, such as identification of distribution, first-order statistics calculation, and receiver operating characteristic curve analysis, was performed.

Results

The average dice index is about 0.6, and average sensitivity and specificity are about 81% and 99%, respectively. The value of sensitivity and dice index are influenced by the number of false positives and false negatives. Because artifacts and infarcts have similar imaging characteristics, it is difficult to completely eliminate the artifacts. The accuracy of localization is nearly 100% as there were only two false-positive and three false-negative slices of all 381 slices. The algorithm takes about 1 minute in the Matlab computing environment to process a volume.

Conclusion

A method to localize and segment the acute brain infarcts is proposed. The method aids the clinician in reducing the time needed to localize and segment the infarcts. The speed of localization and segmentation can be enhanced further by implementing the algorithm in VC++ and using fast algorithms for selection of Gaussian mixture model parameters.

Section snippets

Materials and method

The data consisted of 13 DWI cases. The DWI scans had in-plane resolutions of 1 mm × 1 mm or 1.5 mm × 1.5 mm, and 5-mm slice thickness. The number of slices in DWI scans was from 27 to 33. The matrix size of DWI scans was 256 × 256.

Results

The results of the algorithm were evaluated against the GT marked by the expert. The segmented slices and GT slices of a given volume were overlaid in order to estimate the following quantities:

  • 1

    TPS: slice containing infarct and also identified as infarct slice by the algorithm.

  • 2

    False-positive slice (FPS): slice not having an infarct but identified as infarct slice by the algorithm.

  • 3

    TNS: slice without infarct and also identified as non infarct slice by algorithm.

  • 4

    FNS: slice containing an infarct

Discussion

Speed and accuracy are critical in clinical acute stroke management. In this study, we propose an automatic method for segmentation of acute infarcts, which may speed up the assessment of stroke region for stroke management applications. The method uses techniques such as Gaussian mixture-based modeling, PNN-based selection of slices and use of common features for classification of artifact and infarct, and allows the user to fine tune parameters according to the requirements (eg, selection of

Conclusion

In this study, we have proposed a technique that may assist radiologists and clinicians in speeding up the process of stroke management. The algorithm can be considered as a semiautomatic method as it requires training of PNN and setting some thresholds. The results of localization were high, with only two FPS and three FNS. The average dice index was about 0.6 and average sensitivity and specificity were 81% and 99%, respectively. In the current method, we had fewer FNA voxels, but a higher

Acknowledgments

The authors would like to thank the Biomedical Research Council, Agency for Science, Technology and Research, Singapore. We would also like to thank our colleague Mrs. Aminah Beevi for her assistance in preparing the manuscript.

References (16)

There are more references available in the full text version of this article.

Cited by (32)

  • Applications of artificial intelligence for DWI and PWI data processing in acute ischemic stroke: Current practices and future directions

    2022, Clinical Imaging
    Citation Excerpt :

    For DWI, many automatic segmentation techniques are developed for automatic and real-time lesion volume assessment. They are widely used in DWI imaging and can be divided into three categories: ADC thresholding, 56–59machine learning 60–66and deep learning algorithm using Convolutional Neural Networks (CNNs)67–71. In this review, we describe the PWI/DWI processing pipeline.

  • Automated segmentation of acute stroke lesions using a data-driven anomaly detection on diffusion weighted MRI

    2020, Journal of Neuroscience Methods
    Citation Excerpt :

    Automated approaches could, in theory, eliminate inter-subject variability due to their reproducibility and provide fast results for patient care, as well as facilitate automatic analysis of large datasets (Maier et al., 2015a,b). Recently there has been a growing interest in the automated analysis of neuroradiological images, particularly on AIS lesion segmentation (Maier et al., 2015a,b, Cl’erigues et al., 2019; Perez Malla et al., 2019; Bhanu Prakash et al., 2006; Tsai et al., 2014; Mohd Saad et al., 2011; Menze et al., 2014; Gillebert et al., 2014; Chen et al., 2017; Mah et al., 2014; Boldsen et al., 2018; Stamatakis and Tyler, 2005a; Kamnitsas et al., 2017; Winzeck et al., 2019; Subbanna et al., 2019). The literature on automated and semi-automated stroke lesion segmentation mainly refers to machine learning and neural network methods (Chen et al., 2017; Dou et al., 2016; Mah et al., 2014; Perez Malla et al., 2019; Guerrero et al., 2018; Subbanna et al., 2019; Winzeck et al., 2019; Kamnitsas et al., 2017).

  • White matter hyperintensity and stroke lesion segmentation and differentiation using convolutional neural networks

    2018, NeuroImage: Clinical
    Citation Excerpt :

    Classical supervised machine learning methods such as k-nearest neighbors (kNN) (Anbeek et al., 2004a), Bayesian models (Maillard et al., 2008), support vector machines (SVM) (Lao et al., 2008), and random forests (Geremia et al., 2011) have been well studied in MS segmentation. For stroke lesion segmentation, pattern classification techniques to learn a segmentation function were also employed in Prakash et al. (2006); Maier et al. (2014); Maier et al. (2015b); Maier et al. (2015a). The lesion prediction algorithm (LPA) (Schmidt, 2017), implemented in SPM's LST toolbox, has been shown to produce consistently good performance and in many cases is considered a robust gold standard for this problem.

View all citing articles on Scopus
View full text