Elsevier

Medical Image Analysis

Volume 32, August 2016, Pages 233-242
Medical Image Analysis

A framework for the automatic detection and characterization of brain malformations: Validation on the corpus callosum

https://doi.org/10.1016/j.media.2016.05.001Get rights and content

Highlights

  • We propose a framework for the detection and characterization of brain malformations.

  • We validate the method on the corpus callosum from a pediatric population.

  • We extend the one-class SVM to the multiple kernel framework.

  • We extend the discriminative direction analysis to the multiple kernel framework.

  • We verify the impact on the diagnosis agreement among neuroradiologists.

Abstract

In this paper, we extend the one-class Support Vector Machine (SVM) and the regularized discriminative direction analysis to the Multiple Kernel (MK) framework, providing an effective analysis pipeline for the detection and characterization of brain malformations, in particular those affecting the corpus callosum.

The detection of the brain malformations is currently performed by visual inspection of MRI images, making the diagnostic process sensible to the operator experience and subjectiveness. The method we propose addresses these problems by automatically reproducing the neuroradiologist’s approach. One-class SVMs are appropriate to cope with heterogeneous brain abnormalities that are considered outliers. The MK framework allows to efficiently combine the different geometric features that can be used to describe brain structures. Moreover, the regularized discriminative direction analysis is exploited to highlight the specific malformative patterns for each patient.

We performed two different experiments. Firstly, we tested the proposed method to detect the malformations of the corpus callosum on a 104 subject dataset. Results showed that the proposed pipeline can classify the subjects with an accuracy larger than 90% and that the discriminative direction analysis can highlight a wide range of malformative patterns (e.g., local, diffuse, and complex abnormalities). Secondly, we compared the diagnosis of four neuroradiologists on a dataset of 128 subjects. The diagnosis was performed both in blind condition and using the classifier and the discriminative direction outputs. Results showed that the use of the proposed pipeline as an assisted diagnosis tool improves the inter-subject variability of the diagnosis.

Finally, a graphical representation of the discriminative direction analysis was proposed to enhance the interpretability of the results and provide the neuroradiologist with a tool to fully and clearly characterize the patient malformations at single-subject level.

Introduction

Magnetic Resonance Imaging (MRI) is currently the gold standard technique to study the brain morphology in the medical practice. The detection of brain malformations is performed by trained neuroradiologists, who examine the images and evaluate the presence/absence and the shape of all brain structures. Multiple geometric features (i.e., size, architecture, curvature, thickness, density) are considered and visually inspected to detect possible alterations and malformative patterns. Such approach is useful but suffers from the limitation of being subjective and only qualitative. As a consequence, the diagnosis of subtle malformations is scarcely reproducible among different neuroradiologists and it is usually influenced by the overall clinical pattern of the subject, rather than by the structure shape alone.

Automatic classifiers, such as Support Vector Machine (SVM) (Schölkopf, Smola, 2001, Vapnik, 1998), can be used to support the diagnostic task, providing mathematical tools to describe and classify either the single structure or the whole brain, therefore improving the reproducibility of the diagnosis. In particular, the one-class SVM methods (Schölkopf, Platt, Shawe-Taylor, Smola, Williamson, 2001, Tax, Duin, 2004) have been specifically designed for the outlier detection, i.e., for the identification of objects which do not belong to a specific group. Moreover, Multiple Kernel (MK) methods (Gönen and Alpaydın, 2011) can efficiently combine features from different descriptor functions, hence including all the different aspects usually taken into account by neuroradiologists for a given brain structure. Therefore, SVM methods can be used to automatically reproduce the neuroradiologist’s approach in the diagnosis of morphologic malformations of the brain. This can not only improve the intra-subject reproducibility of the diagnosis, but also the inter-subject agreement in the diagnosis, particularly in the borderline cases.

On the other hand, the main drawback of automatic classifiers is that they are usually developed as “black box” tools that provide the subject classification, but not a description or a list of the aspects that led to the choice. This hinders the diffusion of diagnosis support tools based on automatic classifiers, since the detected malformation should be presented for a validation and its characterization is expected in the medical report.

In Golland et al. (2005), Golland and colleagues introduced the concept of discriminative direction as the direction in the sample space that affects the output classifier, while introducing as little irrelevant changes as possible into the input vector. The discriminative direction tool can be used to highlight the features that caused the assignation of the subject to a specific group, giving the opportunity to not only detect but also describe the critical aspects in the abnormal structure. Moreover, a subsequent evolution, the regularized discriminative direction (Zhou et al., 2009), gives the opportunity to include also a prior knowledge in the analysis. In particular, Zhou and colleagues noticed that when deforming a shape along the original discriminative direction proposed in Golland et al. (2005), spurious shape differences may appear, as a solution may be mathematically acceptable, but not from a physiological point of view. The solution proposed in Zhou et al. (2009) allows to overcome this problem by introducing a physiological constraint directly derived from the training set.

In this work, we propose a fully automatic framework for the detection and characterization of brain malformations. The analysis pipeline exploits the one-class SVM methods embedded in a MK framework to reproduce the neuroradiologist’s approach to the diagnosis problem. Moreover, we apply the regularized discriminative direction analysis to the one-class MK classifier to provide not only the subject classification, but also a description of the malformation. In particular, we verify the effectiveness of the proposed pipeline in the detection of malformations of the corpus callosum. Finally, we study the effectiveness of the proposed pipeline as an assisted diagnosis tool (ADT) comparing the inter-subject agreement rate between the blind-diagnosis and the diagnosis assisted by the classifier. This work expands the results presented in Peruzzo et al. (2014a), where only two-class SVMs and the original formulation of the discriminative direction analysis were applied to the problem and no study was performed to evaluate the impact of the proposed method on the diagnostic process. One-class SVMs are more suitable than standard SVMs to deal with heterogeneous brain abnormalities, as in our case-study. Only healthy subjects are used to estimate the SVM model, while patients are considered as general outliers, not necessarily associated to a particular class.

We focused our analysis on the corpus callosum (CC) that represents the largest commissural structure in the human brain, playing a fundamental role both in motor and cognitive functions. It has a well-defined shape on the midsagittal cerebral plane, appearing as a broad-arched structure, with a variable thickness along its major axis. It is frequently involved in several genetic conditions showing a wide range of malformative patterns (i.e., complete or partial agenesis, hypoplasia, thinning, thickening) Bodensteiner et al. (1994); Kendall (1983); Paul et al. (2007); Van Bon et al. (2008) associated with mental retardation and motor impairment. Alterations in the morphology of the CC can therefore represent a clue of more widespread brain derangement and their identification and quantification is of the utmost importance in the clinical practice.

Section snippets

Feature extraction

The CC feature extraction procedure was implemented on MRI T1-weighted structural images. Images were firstly preprocessed using the FMRIB Software Library (FSL) (Jenkinson et al., 2012). In particular, images were corrected for the magnetic field inhomogeneities, non-brain tissue was removed and the resulting brain images were rigidly registered to the MNI atlas. Finally, a segmentation operation was performed to obtain a probability map for the White Matter (WM) tissue. All preprocessing

Experiment 1

The aim of the first experiment was to validate the proposed pipeline. In particular, we assessed the classification performances of the different classifiers in order to select the best one and then applied the discriminative direction analysis to verify if it could identify the malformative aspect of the CC at subject level.

From the database of pediatric brain MRI reports of the Scientific Institute “Eugenio Medea”, two experienced pediatric neuroradiologists selected 52 subjects with CC

Experiment 1

Following the proposed pipeline, we computed the geometric features for each subject. A visual inspection confirmed that CC masks and features were correctly extracted. Thus, no manual interventions or post-hoc corrections were performed on the automatically extracted features.

The performance of each classifier was evaluated computing several indices: accuracy, sensitivity, specificity, precision and the Receiver Operating Characteristic (ROC) curve, which is characterized by its Area Under the

Discussion

In our work, we applied several kernel based classifiers to a routine neuroradiological problem, the detection of brain malformations from MRI data and their characterization. In particular, we focused on the malformations of the CC, which is an important structure of the brain that can be affected by a wide range of malformative patterns.

We tested six classifiers, three using a standard two-class approach and three using the one-class approach proposed in Tax and Duin (2004). For each of them,

Conclusion

In this study we proposed a fully automatic framework to support the physician in the diagnosis of brain malformations from MRI data, in particular in the CC.

We extracted several shape features of the CC and applied a one-class MK classifier that naturally integrates the different measures, thus reproducing the neuroradiologist’s approach.

We used the regularized discriminative direction analysis to detect the discriminative features for each single subject, thus providing a qualitative and

References (24)

  • B.E. Kendall

    Dysgenesis of the corpus callosum

    Neuroradiology

    (1983)
  • J.-Y. Kwok et al.

    The pre-image problem in kernel methods

    IEEE Trans. Med. Imaging

    (2004)
  • Cited by (8)

    • Huberized one-class support vector machine with truncated loss function in the primal space

      2022, Advances in Engineering Software
      Citation Excerpt :

      It is first proposed by Scholkopf [9]. The OCSVM, which has been used successfully in anomaly detection [10,11], Time-series novelty detection [12], unbalanced classification [13], medical image analysis [14], and network security [15], has been attracting a lot of interest. The model of OCSVM is similar to that of SVM and is a hyperplane-based one-class SVM.

    • Artificial intelligence applications in medical imaging: A review of the medical physics research in Italy

      2021, Physica Medica
      Citation Excerpt :

      In the following sections we present the most representative studies on AI applied to imaging in Italy, grouped according to the clinical use of AI. The most common neuroimaging issues faced with AI tools are early diagnosis, biomarker identification and understanding the mechanism of development of neurodegenerative [99,100–102] and oncological diseases [103,104], mental disorders [101,105–109] and brain malformations [110]. Applications of AI in neuroimaging most commonly takes advantage of MRI, by applying ML to radiomic features ([111]) or directly to voxel intensity [88,90,112,113].

    • Brain Tissue Segmentation Using U-Net-based Convolutional Neural Networks

      2023, RISTI - Revista Iberica de Sistemas e Tecnologias de Informacao
    View all citing articles on Scopus
    View full text