G. Saravana Kumar

Automatic segmentation of vertebral contours from CT images using fuzzy corners

Automatic segmentation of bone in computed tomography (CT) images is critical for the implementation of computer-assisted diagnosis which has increasing potential in the evaluation of various spine disorders. Of the many techniques available for delineating the region of interest (ROI), active contour methods (ACM) are well-established techniques that are used to segment medical images. The initialization for these methods is either through manual intervention or by applying a global threshold, thus making them semi-automatic in nature. The paper presents a methodology for automatic contour initialization in ACM and demonstrates the applicability of the method for medical image segmentation from spinal CT images. Initially, a set of feature markers from the image is extracted to construct an initial contour for the ACM. A fuzzified corner metric, based on image intensity, is proposed to identify the feature markers to be enclosed by the contour. A concave hull based on α shape, is constructed using these fuzzy corners to give the initial contour. The proposed method was evaluated against conventional feature detectors and other initialization methods. The results show the method׳s robust performance in the presence of simulated Gaussian noise levels. The method enables the ACM to efficiently converge to the ground truth segmentation. The reference standard for comparison was the annotated images from a radiologist, and the Dice coefficient and Hausdorff distance measures were used to evaluate the segmentation.

Pull out strength of pedicle screw in normal and osteoporotic cancellous bone models

Lower back pain is treated by bone fusion between adjacent vertebrae typically using instrumentation to immobilize and stabilize the spine till bony fusion takes place. Instrumentation is effected surgically using pedicle screws. However breakage and loosening of pedicle screws from the vertebral body are two main clinical concerns particularly for osteoporotic patients. The objective of the present study is to evaluate the effect of bone density, insertion depth and insertion angle of pedicle screw on the pull out strength and insertion torque using rigid polyurethane foam constructs. Insertion torque was measured during the insertion of the screw and pull-out tests were performed on instrumented pedicle screws and foam construct. These rigid foams represented the range of extremely osteoporotic to normal bone densities. The insertion angle represented the range of angle of pedicle and the insertion depth represented the depth to which these pedicle screws are inserted in the vertebra. The results showed that the pull out strength and insertion torque increased with increase in density and insertion depth (p<0.05) whereas pull out strength decreased with increase in insertion angle. There was significant interaction effect of the various factors on the pull out strength and insertion torque. The findings from the current study suggest that holding power of the pedicle screws can be increased by increasing the purchase depth and reducing the angle of insertion. A statistical model was developed to predict the holding power of the screw which can assist surgeon in pre surgical planning.

Automatic initialization for segmentation of medical images based on active contour

Image segmentation forms the crux of an image processing module. Its main goal is to divide the image into foreground and background with region of interest occupying the former. To achieve quality segmentation particularly in medical imaging, algorithms require user interaction with domain knowledge. Active contours are well established techniques which are prolifically used for segmenting medical images. It falls under the semi automatic method owing to the manual contour initialization. In this paper, we propose a simple, global approach of segmentation using an automatic initialization for active contour methods which has the potential to completely eliminate the human interaction.

Simulation based expert system to predict the deep drawing behaviour of tailor welded blanks

The forming behaviour of tailor welded blanks (TWB) is influenced by sheet thickness ratio, strength ratio and weld conditions in a synergistic fashion. In most of the cases, these parameters deteriorate the forming behaviour of TWB. It is necessary to predict suitable TWB conditions for achieving better stamped product made of welded blanks. This work primarily aims at developing an expert system based on artificial neural network (ANN) model to predict the deep drawing behaviour of TWBs made of steel grade base materials. The important deep drawing characteristics of TWB namely maximum draw depth and weld line profile are predicted within wide range of varied blank and weld conditions. The square cup deep drawing test is simulated in an elastic-plastic finite element code, PAM STAMP 2G®, for generating the required output data for ANN training and validation. The predictions from ANN are encouraging with acceptable prediction errors.

Numerical studies on fluid flow characteristics through different configurations of spiral casing

ABSTRACT The design of optimum spiral casing configuration is a difficult task and a big challenge in the field of turbomachinery. Computational fluid dynamics (CFD) analysis of fluid flow characteristics in a turn around spiral casing plays an important role in its design. The objective in this study is to propose an optimum spiral casing configuration by reducing the total pressure loss and increasing the spiral velocity coefficient and average radial velocity at the exit of spiral casing. For this, three different configurations of spiral casing, viz. accelerated, decelerated and free vortex type, with different aspect ratios (ARs) are numerically simulated. A Eulerian velocity-correction method based on the streamline upwind Petrov–Galerkin (SUPG) finite-element method is employed to solve complete Reynolds-averaged Navier–Stokes (RANS) equations governing fluid flow characteristics. The results show that the average radial velocity along the exit of spiral casings is more for elliptical cross-sectional spiral casings of AR>1 when compared to circular cross-sectional spiral casings. The total pressure loss is found to be at minimum for decelerated spiral casings. In the case of decelerated spiral casings, a further reduction in total pressure loss is obtained with elliptical cross-sections of AR>1. The spiral velocity coefficient is found to be at maximum for decelerated spiral casings with AR>1.

Computing the tensile behaviour of tailor welded blanks made of dual-phase steel by neural network-based expert system

This work primarily aims to develop an expert system based on the artificial neural network (ANN) to predict the tensile behaviour of tailor welded blanks (TWBs) made of dual-phase (DP) 590 steel. The work also aims to compare the predictions by ANN models with empirical models and the size of the training data set of the prediction accuracy of these models. The strain hardening exponent ‘n’ and strength coefficient ‘K’ are predicted. The results obtained from expert system and empirical models are validated by comparing them with the results obtained from finite element simulations and experiments. It is observed that expert system/ANN predictions based on the full factorial design of experiments (DOE) is better than the ANN predictions based on the orthogonal array and predictions based on the empirical models. With the reduced orthogonal training data, ANN model-based predictions are more accurate than the empirical models in most of the test cases taken outside the training range. Inverse models for predicting the TWB parameter combination for good tensile characteristics are also developed and show promising results. The ANN/expert system developed through full factorial DOE is comparable with the experimental results. The ANN/expert system developed through orthogonal DOE is not comparable with the experimental results.

Hybrid intelligent systems for pneumatic sequential circuit design: Hybridizing genetic algorithms and rule-based logic programming

Logic synthesis is an important step in process design and automation. Electronics witnessed tremendous improvements in the reliability of automatically designed circuits with reduced lead time with the application of logic synthesis in circuit design. Industrial process automation systems using hydraulic and pneumatic circuits need logic synthesis to obtain proper sequence of process operations. Deterministic approaches have been used for decades to aid in logic synthesis. These approaches are limited by the volume of data that could be handled because of phenomena like exponential explosion. This paper presents an approach based on hybrid intelligence for logic synthesis for industrial automation systems using pneumatic sequential circuits. The proposed logic synthesis system starts with the sequence of operations as its input and generates the truth table required for the pneumatic circuit for performing the given set of operations. The proposed system hybridizes rule based logic programming and genetic algorithms and is capable of solving considerably difficult problems through an evolutionary strategy. Case studies are presented to demonstrate the capability of the proposed approach for logic design for pneumatics based industrial automation.

Prediction of Tensile and Deep Drawing Behaviour of Aluminium Tailor-Welded Blanks

Tailor-welded blanks (TWB) are blanks with sheets of similar or dissimilar thicknesses, materials, coatings welded in a single plane before forming. Applications of TWB include car door inner panel, deck lids, bumper, side frame rails etc. in automotive sector (Kusuda et al., 1997; Pallet & Lark, 2001). Aluminium TWBs are widely used in automotive industries because of their great benefits in reducing weight and manufacturing costs of automotive components leading to decreased vehicle weight, and reduction in fuel consumption. The general benefits of using TWBs in the automotive sector are: (1) weight reduction and hence savings in fuel consumption, (2) distribution of material thickness and properties resulting in part consolidation which results in cost reduction and better quality, stiffness and tolerances, (3) greater flexibility in component design, (4) re-usage of scrap materials to have new stamped products and, (5) improved corrosion resistance and product quality1. The forming behaviour of TWBs is affected by weld conditions viz., weld properties, weld orientation, weld location, thickness difference and strength difference between the sheets (Bhagwan, Kridli, & Friedman, 2003; Chan, Chan, & Lee, 2003). The weld region in a TWB causes serious concerns in formability because of material discontinuity and additional inhomogeneous property distribution. Above said TWB parameters affect the forming behaviour in a synergistic manner and hence it is difficult to design the TWB conditions that can deliver a good stamped product with similar formability as that of un-welded blank. Designers will be greatly benefited if an expert system is available that can deliver forming behaviour of TWB for varied weld and blank conditions. Artificial neural network (ANN) modelling technique is found to show better prediction of any response variable that is influenced by large number of input parameters. Artificial Neural Networks are relatively crude electronic models based on the neural structure of the brain. The building blocks of the neural networks is the neuron, which are highly interconnected. In the artificial neural networks, the neurons are arranged in layers: an input layer, an output layer, and several hidden layers. The nodes of the input layer receive information as input patterns, and then transform the information through the links to other connected nodes layer by layer to the output nodes. The transformation behavior of the network depends on the structure of the

Influence of morphological variations on cervical spine segmental responses from inertial loading

ABSTRACT Objectives: The objective of this study was to investigate the influence of morphological variations in osteoligamentous lower cervical spinal segment responses under postero-anterior inertial loading. Methods: A parametric finite element model of the C5–C6 spinal segment was used to generate models. Variations in the vertebral body and facet depth (anteroposterior), posterior process length, intervertebral disc height, facet articular process height and slope, segment orientation ranging from lordotic to straight, and segment size were parameterized. These variations included male–female differences. A Latin hypercube sampling method was used to select parameter values for model generation. Forces and moments associated with the inertial loading were applied to the generated model segments. The 7 parameters were grouped as local or global depending on the number of spinal components involved in the shape variation. Four output responses representing overall segmental and soft tissue responses were analyzed for each model variation: response angle of the segment, anterior longitudinal ligament stretch, anterior capsular ligament stretch, and facet joint compression in the posterior region. Pearsons correlation coefficient was used to compute the correlations of these output responses with morphological variations. Results: Fifty models were generated from the parameterized model using a Latin hypercube sampling technique. Variation in response angle among the models was 4° and was most influenced by change in the combined dimension of vertebral body and facet depth, followed by size of the segment. The maximum anterior longitudinal ligament stretch varied between 0.1 and 0.3 and was strongly influenced by the change in the segment orientation. The anterior facet joint region sustained tension, whereas the posterior region sustained compression. For the anterior capsular ligament stretch, the most influential global variation was segment orientation, whereas the most influential local variations were the facet height and facet angle parameters. In the case of posterior facet joint compression, segment orientation was again most influential, whereas among the local variations, the facet angle had the most influence. Conclusion: Shape variations in the intervertebral disc influenced segmental rotation and ligament responses; however, the influence of shape variations in the facet joint was confined to capsular ligament responses. Response angle was most influenced by the vertebral body depth variations, explaining greater segmental rotations in female spines. Straighter spine segments sustained greater posterior facet joint compression, which may offer an explanation for the higher incidence of whiplash-associated disorders among females, who exhibit a straighter cervical spine. The anterior longitudinal ligament stretch was also greater in straighter segments. These findings indicate that the morphological features specific to the anatomy of the female cervical spine may predispose it to injury under inertial loading.

Data Augmentation Techniques for Classifying Vertebral Bodies from MR Images

The article describes the effect of data augmentation on classification systems that are used to differentiate abnormalities in medical images. The imbalance in data leads to bias in classifying the various states. Medical images, in general, are deficit of data and thus augmentation will provide an enriched dataset for the learning systems to identify and differentiate between deformities. We have explored additional data generation by applying affine transformations and the instance based transformation that could result in improving the classification accuracy. We perform experiments on the segmented dataset of vertebral bodies from MR images, by augmenting and classified the same, using Naive Bayes, Radial Basis Function and Random Forest methods. The performance of classifiers was evaluated using the True Positive Rate (TPR) obtained at various thresholds from the ROC curve and the area under ROC curve. For the said application, Random Forest method is found to provide a stable TPR with the augmented dataset compared to the raw dataset.