Identifying central and peripheral nerve fibres with an artificial intelligence approach
Graphical abstract
Introduction
The number of published works related to nerve fibres has progressively increased from the beginning of the 20th century as shown in Fig. 1. A rapid search in PubMed revealed the publication of approximately 130,000 pertinent papers during this period. Almost 40% of these, approximately 50,600 articles, were published in the first 14 years of the 21st century, with most (78%) in the last ten years.
The first studies registered in PubMed using morphometric techniques date back to 1969 [1]. Approximately 1600 articles on morphometric research in nerve fibres have been published in the same period, with approximately 30% of these within the last decade. Among them, we could only locate starting in 1969, approximately 400 studies using morphometric methods associated with electron microscopy. Furthermore, publications related to morphometric and ultrastructural studies have been less abundant, especially those related to optic nerve fibres (73 articles) and cochlear nerve fibres (16 articles).
The morphometric study of nerve fibres is a useful approach to research several subjects related to the nervous system, such as development ([[2], [3]], aging, [[4], [5], [6], [7], [8]]) and pathological conditions both in peripheral [[9], [10], [11], [12]] and central nerve fibres [13].
Several morphometric studies have shown a similar relationship between functional features of nerve fibres and their morphological and morphometric parameters, such as (a) number, density, and diameter of nerve fibres [[14], [15], [16], [17], [18], [19], [20], [21], [22], [23]] (b) in vertebrates, the myelin sheath that encircles large axons determines the fibre conduction velocity. In fact, myelin thickness is related to the speed at which an axon can transmit electrical impulses [[24], [25]], and for this reason myelin sheath characteristics are also of interest [[26], [27], [18], [28], [29], [30], [23]] as well as axonal cytoskeletal components [[2], [31], [32], [33]].
Classically, morphometry was performed manually. However, numerous locations in the nervous system contain exceedingly large numbers of nerve fibres; for example, the rat optic nerve contains more than 100,000 fibres [17] and the human optic nerve more than a million [[34], [35], [36]]. In these cases, manual morphometry is very monotonous, tiring, time-consuming, and predisposed to error [37]. Hence, researchers have been adopting different analysis systems to study the morphological and morphometric features of nerve fibres [[8], [36], [38], [39]] in order to significantly reduce data input and processing times. A variety of sampling schemes claim to be capable of resolving this problem and guarantee the reliability of morphometry [37]. Consequently, there is high motivation for the development of automatic morphometry systems.
The application of automatic image processing in fibre recognition has drawn much attention from the image processing and neurology communities. Morphometry that is entirely automated has certain disadvantages, namely, miss-detection and false positives [37]. This manner of automatization has been previously discussed in the literature in cases where the axon is small, illegible, or irregular and could be undervalued due to low contrast as well as other issues related to the automated identification of contours. [[40], [41], [42]]. Currently, automatic morphometry combining interactive image processing has made significant progress [43] in terms of miss-detection and false positives.
This paper proposes a multi-level classifier architecture to resolve the complexity of automatically identifying whether a nerve fibre belongs to the central or peripherical nervous system (CNS or PNS). In particular, we used two supervised techniques, multilayer perceptron (MLP) and decision trees (DT), and an unsupervised technique, K-means clustering. The supervised method is responsible for distinguishing the origin of the nerve fibres (CNS versus PNS) whereas the unsupervised one performs the division of clusters or groups with similar characteristics either for the CNS or PNS.
Recently, advances in AI methods have made possible the development of expert and decision support systems (DSS) in many different areas, such as business analytics, medical diagnostics, psychology, and environmental science. Specifically, a review of the evolution of medical data analysis from a machine learning perspective [44] indicates how AI methods have been applied in the medical field. A study by [45] proposed a hybrid artificial intelligence (based on a fuzzy rule-based system) to forecast outpatient visits with high accuracy. In [46], a trained artificial neural network (ANN) model was developed to predict the weekly number of infectious diarrhoeas by using meteorological factors as input variables.
Among the several useful classifiers in the AI field, we highlight ANN and DT as those most frequently chosen for the construction of DSS [47]. The goal is normally to establish groups or clusters with similar features from the data. Since there are no references or expected classification and the classification is data driven, the system is unsupervised.
We have previously demonstrated that AI methods are capable of improving the accuracy of the final classification as well as of selecting the best features because very often there are many features to control. For instance, we have experience in classification tasks for male fertility [[48], [49]], urology diagnosis [[50], [51]], and brain ventricles in MR images [52]. In [53], a model to diagnose urological dysfunction is presented. The aim of the study was to correlate the neurological aetiology with the neural centres involved in the two urological phases of voiding and micturition. This previous experience may lead to knowledge discovery regarding databases, data mining, or the process of extracting patterns from large datasets. Nowadays, these techniques are also starting to be used in the field of big data [[54], [55]].
In the present study, the main objective was to develop a classifier architecture based on AI methods that could distinguish different types of nerves and classify two main types of nerve fibres (CN versus PN nerve fibres).
The novelty and the main contribution of this work is the proposal of a multi-level hierarchical classifier architecture, which comprises supervised as well as unsupervised methods. The hierarchical classifier simplifies the complexity of identifying whether a nerve fibre belongs to the CNS or the PNS and their respective characteristics. Moreover, it allows using different classification methods according to the specific semantic level by providing a flexible approach.
In general, a multi-level hierarchical architecture provides the benefit of a classification where the number of variables is reduced from the higher to the lower levels of the system. Hence, the classifiers for lower levels could be simpler due to reduction in the variable domain. However, in the case of a complex problem, it is not always possible to select the variables that allow the specification of a detailed taxonomy. Finally, the multi-level architecture allows for the selection of the most appropriate classification for the study of the problem in terms of resolution level. Therefore, this approach could be used to study other biological problems.
This work was concerned with the most relevant parameters for both optic and cochlear nerve fibres. The number of parameters is often large and therefore, weighing is crucial to choose only key parameters involved in the classification process of nerve fibres. The approach chosen to address this issue involved using different AI methods that are both supervised and unsupervised, specifically DT, MLP, and unsupervised K-means. The general differences between central and peripheral nerve fibres are well known, and a trained pathologist can distinguish between them easily under the microscope. Nevertheless, the advantage of this approach is the possibility of automating the identification procedure. In addition, it allowed us to identify the hierarchy present in the characteristics of each fibre, providing a new interpretation of the evolution and development of the nervous system.
The remaining part of the paper is organized as follows: First we start by defining the materials and methods of the study (samples of the study). Then, we continue with a brief description of the AI methods used in this paper: MLP, DT, and K-means. Then we proceed by describing the design of our proposed architecture and the experiments carried out in the Results section: the available data as well as a detailed explanation of the different values of our database. Finally, we describe the subsequent testing carried out to analyse the results and draw relevant conclusions.
Section snippets
Architecture
A generic machine learning approach is able to extract a model from data to predict or classify new inputs. Hence, a classifier could be designed using machine learning techniques to learn the type of fibre characteristics irrespective of the level of the taxonomy to be applied (i.e. the fibre belongs to the CNS or the PNS, to a cochlear or sciatic nerve, to a sensory or motor fibre, and others). The diagram in Fig. 2 shows the procedure followed to classify the different fibres in the nerves.
Computational experiments
In this section, we describe the experiments carried out in order to validate the proposed multi-level architecture, and afterwards, we describe the results obtained. First, we conducted a supervised experimentation to classify between the optic and cochlear nerves as the first level of the architecture. Second, unsupervised experimentation was performed to distinguish the types of fibres within each type of nerve as the second level of the architecture.
Discussion
In this study, we adopted a bio-inspired multi-level classifier architecture of nerve fibres as a novel architecture approach for the automation of image analysis in clinical diagnosis. As a first step, we compared the accuracy of several artificial intelligence methods. Due to the complexity of biological systems, and particularly of the nervous system, it was not been possible to accomplish this task directly. The biological system of nerve fibres makes it necessary to develop a multi-level
Acknowledgments
This work was partially supported by the following grants: The Office of the Vice Chancellor for Research, Development, and Innovation, University of Alicante, Spain, (Grant Vigrob-137 to JDJ); the Chair of Reproductive Medicine, University of Alicante-Bernabeu Institute of Reproductive medicine (Grant 4-12I to JDJ). The Spanish Ministry of Economy and Competitiveness (MINECO/FEDER) under the granted Project SEQUOIA-UA (Management requirements and methodology for Big Data analytics)
References (77)
- et al.
Age changes in the human oculomotor nerve – a stereological study
Ann. Anat. – Anat. Anzeiger
(2009) - et al.
A morphometric study on the longitudinal and lateral symmetry of the sural nerve in mature and aging female rats
Brain Res.
(2008) - et al.
Age related optic nerve axonal loss in adult Brown Norway rats
Exp. Eye Res.
(2005) - et al.
The effects of resveratrol on chronic constriction injury of sciatic nerve in rats
Neurosci. Lett.
(2014) - et al.
Axon number and size distribution in the developing feline inferior alveolar nerve
J. Neurol. Sci.
(1982) - et al.
Axon types classified by morphometric and multivariate analysis in the rat optic nerve
Brain Res.
(1992) - et al.
Myelin sheath thickness and internodal length of nerve fibres in the developing feline inferior alveolar nerve
J. Neurol. Sci.
(1982) - et al.
Internodal myelin volume and axon surface area
J. Neurol. Sci.
(1982) - et al.
A useful programme in BASIC for axonal morphometry with introduction of new cytoskeletal parameters
J. Neurosci. Methods
(1991) - et al.
An electron microscopic analysis of the optic nerve of the eastern chipmunk (Tamias sibiricus asiaticus): Total fiber count and retinotopic organisation
Vis. Res.
(1987)
Automatic identification and morphometry of optic nerve fibers in electron microscopy images
Comput. Med. Imaging Graph.
Histomorphometry of the optic nerves of normal dogs and dogs with hereditary glaucoma
Exp. Eye Res.
Automated morphometric analysis in peripheral neuropathies
Comput. Biol. Med.
Machine learning for medical diagnosis: history, state of the art and perspective
Artif. Intell. Med.
Developing a hybrid artificial intelligence model for outpatient visits forecasting in hospitals
Appl. Soft Comput.
Artificial neural networks for infectious diarrhea prediction using meteorological factors in Shanghai (China)
Appl. Soft Comput.
Predicting seminal quality with artificial intelligence methods
Expert Syst. Appl.
Application of artificial neural networks in the diagnosis of urological dysfunctions
Expert Syst. Appl.
Using support vector machines in diagnoses of urological dysfunctions
Expert Syst. Appl.
3D reconstruction of medical images from slices automatically landmarked with growing neural models
Neurocomputing
Modelling of urological dysfunctions with neurological etiology by means of their centres involved
Appl. Soft Comput.
Using Kaplan-Meier analysis together with decision tree methods C&RT, CHAID, QUEST, C4.5 and ID3) in determining recurrence-free survival of breast cancer patients
Expert Syst. Appl.
Application of decision tree technology for image classification using remote sensing data
Agric. Syst.
Data clustering: 50 years beyond K-means
Pattern Recogn. Lett.
Neural network adaptation process effectiveness dependent of constant training data availability
Neurocomputing
Utilization of discretization method on the diagnosis of optic nerve disease
Comput. Methods Prog. Biomed.
Morphometrisch-ultrastrukturelle Untersuchungen am Hypophysenhinterlappen der Ratte nach Dehydratation
Zeitschrift für Zellforschung und Mikroskopische Anatomie
Postnatal development of microtubules and neurofilaments in the rat optic nerve: a quantitative study
J. Comp. Neurol.
Visual experience during postnatal development determines the size of optic nerve axons
NeuroReport
Comparison of morphometric aspects of light and electron microscopy of the hypoglossal nerve between young and aged male Wistar rats
Cell J.
Morphometric evaluation of the aging process in various human nerve fibers
Okajimas Folia Anatomica Japonica
Renal nerve ultrastructural alterations in short term and long term experimental diabetes
BMC Neurosci.
The relationship between Bell's palsy and morphometric aspects of the facial nerve
Eur. Arch. Oto-Rhino-Laryngol.
Rodent models of chemotherapy-induced peripheral neuropathy
ILAR J.
Automated quantification of optic nerve axons in primate glaucomatous and normal eyes-method and comparison to semi-automated manual quantification
Investig. Opthalmol. Vis. Sci.
Axon diameters in relation to the spike dimensions and the conduction velocity in mammalian a fibers
Am. J. Physiol.
Conduction velocity and diameter of nerve fibers
Am. J. Physiol.
Hypertrophy of nerve fibers to functionally overloaded muscles
J. Comp. Neurol.
Cited by (7)
Three-stage intelligent support of clinical decision making for higher trust, validity, and explainability
2022, Journal of Biomedical InformaticsCitation Excerpt :We think that the distinction that the declaration provides is good for classification, however, clinical practice shows that more and more hybrid systems are being used to support clinical decisions [29–31]. Data-driven methods employ large amounts of empirical data processed with statistical machine learning methods to extract patterns [32–40]. The need for validity and explainability is currently recognized as an essential problem to be solved by researchers [29].
Machine learning models for credit analysis improvements: Predicting low-income families’ default
2019, Applied Soft Computing JournalCitation Excerpt :Feng et al. [6] also tested different techniques for credit databases and reinforce the use of ML for unbalanced data, which occurs in our database. In addition, machine-learning literature has been generating very interesting results based on standard models, such as the studies published by Barboza et al. [40], Chang et al. [32], Gil et al. [44], Mathew et al. [45], Óskarsdóttir et al. [35]. Not only new theoretical and computational models but also real-world applications of machine learning should be discussed.
Integrated Evolving Spiking Neural Network and Feature Extraction Methods for Scoliosis Classification
2022, Computers, Materials and ContinuaMulti-step clustering approach of myelinated nerve fibers in experimental neuromorphology
2021, International Journal of Ambient Computing and IntelligenceA machine learning and integration based architecture for cognitive disorder detection used for early autism screening
2020, Electronics (Switzerland)