Elsevier

Journal of Clinical Neuroscience

Volume 78, August 2020, Pages 175-180
Journal of Clinical Neuroscience

Clinical study
Ependymoma and pilocytic astrocytoma: Differentiation using radiomics approach based on machine learning

https://doi.org/10.1016/j.jocn.2020.04.080Get rights and content

Highlights

  • A radiomics approach is applied for pediatric posterior fossa tumor differentiation.

  • 300 multimodal features are extracted to describe the statistics of the MRIs.

  • Machine learning methods are combined for effective assisted clinical diagnosis.

Abstract

Mandatory accurate and specific diagnosis demands have brought about increased challenges for radiologists in pediatric posterior fossa tumor prediction and prognosis. With the development of high-performance computing and machine learning technologies, radiomics provides increasing opportunities for clinical decision-making. Several studies have applied radiomics as a decision support tool in intracranial tumors differentiation. Here we seek to achieve preoperative differentiation between ependymoma (EP) and pilocytic astrocytoma (PA) using radiomics analysis method based on machine learning. A total of 135 Magnetic Resonance Imaging (MRI) slices are divided into training sets and validation sets. Three kinds of radiomics features, including Gabor transform, texture and wavelet transform based ones are used to obtain 300 multimodal features. Kruskal–Wallis test score (KWT) and support vector machines (SVM) are applied for feature selection and tumor differentiation. The performance is investigated via accuracy, sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) analysis. Results show that the accuracy, sensitivity, specificity, and AUC of the selected feature set are 0.8775, 0.9292, 0.8000, and 0.8646 respectively, having no significant differences compared with the overall feature set. For different types of features, texture features yield the best differentiation performance and the significance analysis results are consistent with this. Our study demonstrates texture features perform better than the other features. The radiomics approach based on machine learning is efficient for pediatric posterior fossa tumors differentiation and could enhance the application of radiomics methods for assisted clinical diagnosis.

Introduction

As the second most common neoplasm and the most common solid neoplasm in children, central nervous system (CNS) tumors cause more deaths than any other childhood malignancies [1], [2], [3]. About fifty percent of pediatric intracranial CNS tumors occur at the posterior fossa, in which ependymoma (EP) and pilocytic astrocytoma (PA) are two typical kinds [1], [4], [5]. EPs are highly cellular tumors and roughly seventy percent of them occur in the posterior fossa [6], [7]. PAs are the most frequent low-grade tumors and have a predilection site in the posterior fossa, too [7]. The similar predilection pathogenic sites lead to a challenging differentiation between EPs and PAs. On the other hand, because exact differentiation closely correlates with the patient's prognosis and different treatment approaches are acquired for various tumors [8], mandatory accurate and specific diagnoses have brought about increased challenges.

Since the introduction of magnetic resonance imaging (MRI) in the early 1990 s, it has become one of the key imaging techniques used for the visualization and management of pediatric brain tumors [3], [9]. MRI has the advantage of noninvasive sampling and so it has been recognized as the best tool to assess the structure of the posterior fossa tumors [10]. However, MRI is highly sensitive for primary lesion detection and different tumors always demonstrate unclear differences in visual appearance [3]. Thus, tumor diagnosis based on visual inspection of medical images may cause some problems, including wrong diagnoses by radiologists with less experience, heavy workload brought to the radiologists and so on.

The leaps of artificial intelligence (AI) have expanded its application in pediatrics and the AI-based decision support is founded on machine learning (ML) [11], [12], [13]. Automatic image analysis and machine learning techniques can help to support clinical decision-making, serving to reduce the variability caused by human visual perception [14]. With high-performance computing, extracting abundant quantitative features from medical images is no longer difficult now. Recently, radiomics is designed to develop decision support tools and radiomics based quantitative analyses have been investigated regarding their predictive/prognostic value in a variety of diseases and they can help to reflect underlying pathophysiology [15]. Multifarious radiologic features ranging from histogram parameters to spatial interactions between intensity levels to textural heterogeneity measures and to transformation based higher order features have been brought up [16], [17], [18], [19], [20], [21].

The application of radiomics analysis in intracranial tumors can be traced back to 1993. Lerski et al. firstly reported an MR imaging texture analysis procedure for identifying tumor constituents in which a four-layer hierarchical decision tree was used. First-order histogram, gradient, and second-order gray-level co-occurrence matrix (GLCM) features were extracted [22]. Ahmed et al. investigated and compared the efficacy of texture features in posterior fossa tumor segmentation. Feature selection based on Kullback-Leibler divergence (KLD) and expectation maximization framework were used [23]. Orphanidou-Vlachou et al. applied texture analysis to produce 279 features to classify the cases, using machine learning techniques including principal component analysis, linear discriminant analysis, and probabilistic neural network [24]. Rodriguez et al. improved discrimination of tumor types using support vector machine (SVM) classifiers on quantitative radiologic MRI features [25]. Spiteri et al. extracted features describing the first- and second-order statistics for each image to identify quantitative features of posterior fossa syndrome, in which feature selection techniques and SVM were used [26]. Fetit et al. carried out texture analysis to extract MRI features and supervised classification algorithms were compared in the classification of tumor types [3]. These studies have promoted the application of radiomics in the diagnosis of pediatric posterior fossa tumors.

In the present work, we differentiated EP and PA using radiomics approach based on machine learning, in which a total of 300 multimodal radiomics features, including texture, Gabor transform, and wavelet transform based ones, were extracted. Kruskal–Wallis test score (KWT) and SVM were applied for feature selection and tumor differentiation. Further analyses including the performance of different types of radiomics features and the related significance analysis were also carried out.

Section snippets

Patients and datasets

In this study, real preoperative MRI dataset of 45 patients (mean age: 7 years; age range: 0–14 years) were enrolled. All of the patients were diagnosed as posterior fossa tumors with histopathological evidence. 135 MRI slices were evaluated retrospectively: 81 slices had histologically verified EP and 54 slices had PA. This study had been conducted with the consent of the guardian of the patients and this study was approved and reviewed by the Ethics Committee of the first Affiliated Hospital

Results

To investigate the differentiation performance using radiomics approach based on machine learning, 300 radiomics features were extracted from the segmented tumor regions of 135 MRI slices belonging to two kinds of pediatric posterior fossa tumors. KWT feature selection and SVM classification training were carried out using the training set (n = 95), whereas the validation set (n = 40) was used to assess the performance. The workflow of image processing and machine learning is shown in Fig. 1.

Discussion

The diagnosis of tumors by human experts may be influenced by experience and the variability among different patients. Radiomics analysis based on machine learning is the most novel approach used to alleviate this problem by capturing a large amount of information that human vision cannot detect. A series of studies have demonstrated that this approach can provide better diagnostic results than human experts [13], [17], [20], [29]. Therefore, this study focuses on the application of machine

Ethical statement

The research is done following all the Ethical Principles for Medical Research Involving Human Subjects outlined in the Declaration of Helsinki.

Acknowledgements

This work is supported by the National Natural Science Foundation of China, Grant U1304602 and 61673353. The MRIs used in this paper are provided and marked by clinical imaging specialists of the Magnetic Resonance Department, the first Affiliated Hospital of Zhengzhou University and the authors express their gratitude to them.

References (38)

  • L.A. Brandão et al.

    Posterior fossa tumors

    Neuroimag Clin N Am

    (2017)
  • R.A. Lerski et al.

    MR image texture analysis-an approach to tissue characterization

    Magn Reson Imaging

    (1993)
  • M. Idrissa et al.

    Texture classification using Gabor filters

    Pattern Recogn Lett

    (2002)
  • H. Suk et al.

    Deep ensemble learning of sparse regression models for brain disease diagnosis

    Med Image Anal

    (2017)
  • A. Poretti et al.

    Neuroimaging of pediatric posterior fossa tumors including review of the literature

    J Magn Reson Imaging

    (2012)
  • E. Hanzlik et al.

    A systematic review of neuropsychological outcomes following posterior fossa tumor surgery in children

    Childs Nerv Syst

    (2015)
  • A.E. Fetit et al.

    Three-dimensional textural features of conventional MRI improve diagnostic classification of childhood brain tumours

    Nmr Biomed

    (2015)
  • S. Zitouni et al.

    Apparent diffusion coefficient in differentiation of pediatric posterior fossa tumors

    Jpn J Radiol

    (2017)
  • N.D. Sabin et al.

    Quantitative imaging analysis of posterior fossa ependymoma location in children

    Childs Nerv Syst

    (2016)
  • W. Monika

    Imaging and diagnosis in pediatric brain tumor studies

    (2017)
  • S. Higano et al.

    Malignant astrocytic tumors: clinical importance of apparent diffusion coefficient in prediction of grade and prognosis

    Radiology

    (2006)
  • R. Awa et al.

    Neuroimaging diagnosis of pineal region tumors-quest for pathognomonic finding of germinoma

    Neuroradiology

    (2014)
  • C. Raybaud et al.

    Posterior fossa tumors in children: developmental anatomy and diagnostic imaging

    Childs Nerv Syst

    (2015)
  • P. Kokol et al.

    Artificial intelligence and pediatrics: a synthetic mini review

    Pediatr Dimensions

    (2017)
  • L. Shu et al.

    Application of artificial intelligence in pediatrics: past, present and future

    World J Pediatr

    (2019)
  • P. Yin et al.

    Comparison of radiomics machine-learning classifiers and feature selection for differentiation of sacral chordoma and sacral giant cell tumour based on 3D computed tomography features

    Eur Radiol

    (2019)
  • A. Kassner et al.

    Texture analysis: a review of neurologic MR imaging applications

    Am J Neuroradiol

    (2010)
  • R.J. Gillies et al.

    Radiomics: images are more than pictures, they are data

    Radiology

    (2016)
  • Ford J, Dogan N, Young L, Yang F. Quantitative radiomics: impact of pulse sequence parameter selection on MRI-based...
  • Cited by (0)

    1

    Mengmeng Li and Haofeng Wang contribute equally to this work. (H. Wan and Z. Shang)

    View full text