Computer aided detection of spina bifida using nearest neighbor classification with curvature scale space features of fetal skulls extracted from ultrasound images
Introduction
With the advent of computer science and automatization of many tasks in many fields, realizing computer aided diagnosis (CAD) systems has been an area of interest in biomedicine. CAD usually refers to procedures in medicine to help specialists in the interpretation of medical images for decision-making. Typically, CAD is an interdisciplinary framework consisting of radiological and digital image processing combined with machine learning. Imaging techniques of X-ray, magnetic resonance (MRI), computed tomography (CT) and ultrasound (US) are the main sources of data (i.e. images) to be used as inputs to CAD systems. The structures of interest (i.e. regions of interest (ROI)) in images are usually analyzed to detect the presence of conspicuous structures which help identify particular diseases. CAD systems may only be intended to act as supporting agents and not substitute doctors who are responsible for the final interpretation of images in an absolute and ethical sense. Examples of CAD applications include the diagnosis of various cancer types (e.g. lung, colon, breast) [8], [20], [28], coronary artery disease [2], mammographic masses [7], peripheral soft tissue masses [6], and more.
In the CAD application subject to this paper, our aim is the automatized detection of the common neural pathology called spina bifida (open/split spine) from ultrasound (US) images of fetal skulls acquired in prenatal terms. As the name implies, spina bifida is one of a group of defects known as neural tube defects (NTD) related to the spine and spinal cord. In neural development of embryos, a tissue called the neural plate folds and forms a tube, which then folds into the spinal cord. Incorrect folding of the neural plate causes the spina bifida defect [9], whose result is an abnormally formed section of the spinal column. Abnormality occurs at some vertebral column location, referred to as the lesion level. People suffering from spina bifida experience loss of body control below the lesion. The higher the lesion level (the more cranial or the closer to the brain), the more severe the defect. Loss of body control may appear as problems in bladder control, sensation loss and paralysis. Fig. 1 shows the sagittal section of a defective spine of a fetus viewed with US. The prevalence of spina bifida is 1–2 cases per 1000 births worldwide and the incidence is observed to vary up to 3–4 cases per 1000 in some populations. Although what causes the injury is not well known; the consumption of folic acid, a type of vitamin B, by pregnant women shows to prevent up to 70% of neural tube defects including spina bifida.
US examination is a convenient tool to discover neural tube defects in the prenatal stage. Detection of the defect before birth leads of careful planning and effective remedy. Surgical treatment after birth and fetal surgery may be possible, however, most pregnancies with neural tube defects are terminated because of the poor future quality of life of newborns having such defects. Observing the spine itself for detection is not a necessity because fetal heads contain markers indicating the presence of spina bifida. The so-called lemon sign [22] that appears when the frontal bones of the skull look flattened and inwardly bent, is a very typical marker. Fig. 2 shows the transcerebellar section of a malformed skull of a defective fetus with lemon sign.
Automatically deciding whether a fetus is associated with the spina bifida defect (i.e. label 1) or not (i.e. label 0) is obviously a classification problem. Robust classification, from a general viewpoint, involves a number of subtasks such as preprocessing inputs to represent them by means of descriptive features and treating the extracted features with appropriate machine learning methods to assign patterns/classes/labels to them with objectives of optimizing pre-defined performance criteria. Each subtask may further be composed of other subtasks. Specific problems; varying in aspects of input and feature modalities, the number of samples available for learning, the distribution of samples belonging to different classes, the primary objective of classification and how success is perceived; require specific actions to be undertaken. The overall success depends on the selection of algorithms used in each subtask with careful consideration of each detail. Assessing classification performance and generalizing outcomes is also a matter of data sets and methodology employed throughout experiments [10], [11], [12], [30], [31].
Since spina bifida and lemon sign are tightly-coupled and the latter is associated with outer lines of skulls, features extracted from skull contours are supposed to be appropriate for classification. Features along a contour can be captured at sampled points and their combination at all points in a well-enough resolution provide a description of the whole contour. The naturally-arising concept is curvature of points that characterizes how the contour (or curve) bends at the corresponding points. In the curvature scale space (CSS) representation of curves [26], [27], the description is performed by features invariant under translation, rotation and scale. The properties of invariance are definitely very valuable in robust classification because, as applicable to the classification of skull contours, identical shapes must output identical classification results no matter how they are translated, oriented and scaled. One property of the CSS representation is that features for different contours may be of different sizes and classifiers that work on the principle of constructed models using feature vectors of equal size are impractical. In the availability of a scheme that can estimate how similar (or different) two curves are, the lazy k-nearest neighbor (kNN) can be used to compare curves to others with known labels and decide based on similarity. Fortunately, a matching procedure that measures the distance between CSS features of two curves [1] was designed justifying the use of nearest neighbor classifiers.
In the scheme that we propose in this paper, contour lines of fetal skulls are presented to the system and invariant features are computed at multiple scales. The CSS representation [1], [24], [25], [26], [27] of contours produces a map of sampled points with curvatures of zero value at a number of scales (levels of detail (LOD)). As a result, the CSS image of the associated contour from which features can be extracted is obtained. Parametric representations for curves allow analytic solutions in curvature computation. With curves in digital images, such functional forms are not available and techniques exploiting concepts related to the definition of curvature can be utilized. We use differential turning angles (dTA) [16], [17], [18], [19] and their scalograms (dTASS – differential turning angle scale space) to obtain points of zero curvature (zero-crossings) at all scales of consideration. Different scales of a contour correspond to its smoothed versions, each with a different standard deviation of a Gaussian kernel used for smoothing. CSS images are fed to the CSS matching algorithm of Abbasi et al. [1] in a pairwise fashion to output a matching (similarity or distance) score for the two CSS images (hence for the two contours). Classification is performed by kNN.
Fig. 3 shows the block diagram of the proposed CAD system which also includes the module for segmenting input US images to isolate fetal skulls from their surroundings. The work of this article assumes that segmentation, which is a challenging problem on its own right, has been solved and skull shapes (i.e. contours) are ready to be processed by the other modules of the system. Fig. 4, Fig. 5 display four typical US images of fetal skulls viewed as transcerebellar sections and their associated shapes (contours), respectively.
The US image data used in this paper are obtained from the Obstetric and Gynecology Clinics of the Medical Faculties of Istanbul University and Trakya University. Their use for research has been ethically approved by the committees of the mentioned medical centers, provided that patient IDs are kept confidential.
As a matter of fact, assessing classifier performance is not always trivial. This is especially true in medicine when available data sets used in training and to report results have considerably few elements and those that can be collected comprise very unbalanced sets. This problem is addressed by research on methods to deal with rare classes and cases [11], [12], [31]. Sampling the rare class [31] is the most popular technique used to overcome problems of unbalanced data sets. Receiver operating characteristics (ROC) curves and precision-recall plots [10] are generally considered to be good indicators of classification performance. The Area Under the ROC Curve (AUC) metric [5], being the area between ROC curves of classifiers and bounding axes defined by true positive (TP) and false positive (FP) rates, is a metric that quantifies the ROC curve as a whole (i.e. a whole-curve metric) and assigns a numerical value to the performance of classifiers. Point metrics such as F-measure, GMPR, etc. derived from precision and recall defined with respect to the rare class are also in common use.
The remainder of the article is organized as follows: In Section 2, we present our previous work on computerized detection of spina bifida. Section 3 considers curvature scale space representation, CSS image computation using dTASS scalograms, features derived from CSS images and how they can be enhanced. A matching procedure that computes a matching score of CSS images (with its two variants) is described in Section 4. In Section 5, we introduce the data set used in the experiments, detail the classification algorithm and present results obtained with a number of settings. These settings are related to different ways of utilizing data sets, feature sets and computation of matching scores. The presentation takes place within a comparative perspective outlined by the settings. In addition, comparisons with other studies [13], [14] are provided in this section. Section 6 presents discussions on various aspects of the utilized scheme and Section 7 concludes the article.
Section snippets
Related work
We have previously attempted to handle the problem of spina bifida detection with different methodologies [13], [14]. In these works, fetal US images of transcerebellar skulls have been used as inputs and candidate solutions have been realized.
In the work of [13], we fit elliptical models to localize skull boundaries and detect contour segments of fetal skulls in neighborhoods of the ellipses via search. The discontinuous nature of skull boundaries in US images results in the detection of
Curvature scale space representation
There are various approaches that can be employed to represent and describe 2D planar shapes. Each of these can be either contour-based or region-based depending on whether shape features are extracted from only the contour of the shape or from the entire shape region. Representation methods are also classified as global or structural. Global methods are those where a shape is represented as a whole, whereas structural methods describe the shape as segments/sections. Further distinguishing is
CSS matching
Two CSS images that are representations of two different contours can be compared in order to produce a matching score corresponding to the similarity of the two contours or a difference figure to indicate how different from each other the contours are. In the matching procedure, one need not deal with normalizations to compensate for translation, rotation and scale invariance because the CSS representation itself has the invariance property with respect to all these transformations.
The
Classification and results
We use CSS images of normalized [3] skull contours extracted from fetal transcerebellar US images as inputs to classification. The classifier uses the k-nearest neighbor (kNN) method with a fixed value of k (i.e. k = 20). For an illustration of the involved features, we consider the CSS images in Fig. 18 of the normalized skull contours in Fig. 17, which are associated with the fetal skulls viewed in Fig. 16. The CSS image of the first contour (extracted from the first US image) contains 4 arcs
Discussion
We have proposed a new classification approach that can be used in various CAD systems where object contours possess indicative signs to help diagnosis. Our methodology has attacked the specific problem of the spina bifida neural tube defect, most commonly diagnosed using transcerebellar skull images of fetuses acquired via ultrasound viewing.
Instead of defining contour features explicitly, that is instead of defining feature vectors of equal size to identify contours, and running parametric
Conclusion
Spina bifida is among the most common birth defects that may seriously affect the quality of life of individuals after birth. There is no cure for nerve damage caused by the defect. If no action is taken before birth, what can be done to prevent further damage and infections is neurosurgical operation to close the opening on the back by putting the spinal cord and nerve roots back inside the spine and covering with nervous membrane (i.e. meninges). A shunt may also be surgically installed to
Acknowledgements
We would like to express our sincere thanks to the Obstetrics and Gynecology Departments of Medical Faculties of Trakya University and Istanbul University. We are grateful to İbrahim Kalelioğlu of Istanbul University for his help in ultrasound image collection. This work is also being supported by the Scientific Research Projects fund (BAP 14A01P2) of Boğaziçi University and the Turkish Ministry of Development under the TAM Project, Number 2007K120610.
References (33)
The use of the area under the roc curve in the evaluation of machine learning algorithms
Pattern Recogn.
(1997)An introduction to ROC analysis
Pattern Recogn. Lett. – Special issue: ROC Analysis in Pattern Recognition
(2006)- et al.
Multiscale contour description for pattern recognition
Pattern Recogn. Lett.
(2006) - et al.
A review of computer-aided diagnosis of breast cancer: toward the detection of subtle signs
J. Franklin Inst. – Special Issue: Medical Applications of Signal Processing Part I
(2007) - et al.
Review of shape representation and description techniques
Pattern Recogn.
(2004) - et al.
Curvature scale space image in shape similarity retrieval
Multimedia Syst.
(1999) - et al.
Automated computer-aided stenosis detection at coronary CT angiography: initial experience
Eur. Radiol.
(2010) - et al.
Affine-invariant curve normalization for object shape representation classification and retrieval
Machine Vis. Appl.
(2001) - F. Bergeaud, S.G. Mallat, Matching pursuit of images, in: Proceedings of the IEEE International Conference on Image...
- C-Y. Chen, H.K. Chang, Y-T. Kuo, H-J. Chiou, P2D-10 Computer-aided diagnosis of peripheral soft tissue tumors using...
Detection of masses in mammograms via statistically based enhancement, multilevel thresholding, segmentation and region selection
Computer. Med. Imag. Graph.
Computer-aided diagnosis systems for lung cancer: challenges and methodologies
Int. J. Biomed. Imag.
Cited by (10)
An optimized modified faster region convolutional neural network for spina bifida identification from ultrasound images
2023, Biomedical Signal Processing and ControlA fire-controlled MSPCNN and its applications for image processing
2021, NeurocomputingA novel evolutionary method for spine detection in ultrasound samples of spina bifida cases
2021, Computer Methods and Programs in BiomedicineCitation Excerpt :Konur et al. have mostly focused on cranial morphology rather than the spinal axis, which may provide indirect clues about the existence of the pathology. Their algorithm is based on classifying the space features of fetal skull contours using nearest neighbor classification method [2]. This approach also utilizes ultrasound samples, but its efficiency is limited to open SB cases where the skull morphology changes significantly.
Computerized detection of spina bifida using SVM with Zernike moments of fetal skulls in ultrasound screening
2018, Biomedical Signal Processing and ControlCitation Excerpt :The drawbacks include adhoc approaches for segmenting skulls [7,8] and spines [8] in addition to the uncertainty in the number of Zernike moments that should be used [8]. The latest two efforts [9,10] have benefited from the curvature scale space (CSS) [11–14] representation of curves to extract skull features. CSS provides descriptions of contours at multiple scales (levels of detail (LOD)).
Semi-automatic heuristic segmentation of fetal skull images
2023, Journal of the Faculty of Engineering and Architecture of Gazi UniversityEffective Detection of Exomphalos in Fetal Ultrasound Images Using Canny edge detector
2022, Proceedings of the 2022 International Conference on Innovative Computing, Intelligent Communication and Smart Electrical Systems, ICSES 2022