Nowadays, computer assisted techniques is becoming a more promising approach in medical analysis. Developing a robust and efficient approach capable of robustly determining the location of the spine and the contour of the vertebrae are essential, besides assessing vertebral abnormality. The feasibility of computer assisted techniques for the segmentation of vertebra bodies in spine x-ray images has been of great interest to biomedical researches. In general, the quality of computer assisted segmentation is affected by three important factors. First factor is the region of interest (ROI) where the vertebra has a wide ranges of shape, size and orientation, where the second factor is the image quality where the x-ray images are usually of poor quality, segmentation methods often confuse tissue and vertebra boundaries and third is image size and resolution where the cervical or lumbar spine x-ray images too large. However, fully automated segmentation of spine x-ray images is a very challenging problem.
Medical Image Indexing And Retrieval
Medical Images Characteristic
Medical images are multi-modal, where each modality reveals anatomical and /or functional information of different body parts and as its own set of requirement such as file format resolution, dimensionality, and image acquisition and production techniques. It may be required to generate images with different modalities such as (CT, MRI and x-ray) for proper diagnosis for the same patient. DICOM is the common file format for medical images which contains some additional information regarding image modality, acquisition device and patient identification wit raw image data. Two dimensional DICOM image have size much larger tan other general image format (JPEG, GIF and TIFF etc). Higher dimensionality is another property of medical images such as 3-D or 4-D and this type of biomedical images are often organized as collections of these 2-D slices, where the added dimension may be tackiness of slice. Medical images such PET, SPECT and fMRI contain functional information and these need special attention and requirement such enhancement and filtering when performing images analysis task in medical domain.
Investigation Of Segmentation Approach
The segmentation of vertebral body helps to detect and find the contour of vertebrae to categorize vertebrae shape, facilitate shape indexing and recognition in the computer assisted system. Many segmentation algorithms and their result was published using x-ray images to achieves satisfactory shape contour segmentation, this methods have been developed in the past to segment the cervical and the lumbar vertebrae and can be listed as follow,
Active shape contour segmentation (ACS) which is classical snake model combined with and initial contour developed from a prior information and reach the constraint of the shape contour , the algorithm minimizes an objective function by seeking a contour wit maximized gradients along orthogonal curves to the contour and minimized the contour length. Active contour or snakes are computer generated curves tat move within images to find object boundaries. The 3D version is often knows as actives surfaces. ACS models are often used in computer vision and image analysis to detect and locate objects.
Tagare`s ACS is implemented as a semi-automatic witch constrains the solution to lie on grid between an inner contour inside the template and an counter contour outside the template both inner and outer contour are used defined. Other active contour models have been implemented including the traditional snake, the snake using a local minimization algorithm, template deformation along the orthogonal curves of counter and the snake using gradient vector flow. All this algorithms behave very poor of snake, small and discontinuous edges inside, outside and along the shape boundary because of the low image quality such noisy
x-ray images as result the snake often misses the actual boundary of vertebra, were using the grading vector flow performs worst because this method capture a large features at the actual boundary as well the areas wit false edges. Generally speaking, using snakes need intervention and some prior knowledge of the contour to finer segmentation. In other and inaccurate position, scale and rotation of the initial template will result badly segmentation.
A customized version of active shape models (ASM) was implemented by Zamora for the segmentation of the cervical and lumbar vertebrae which was originally formulated by Cootes. The ASM is an optimization algorithm that requires and initial estimation of the position, orientation and scale of the vertebra were this estimated initialization is termed as the pose of vertebra. ASM model considered as a kind of statistical method of object localization based on Eigen-analysis. After performing principal component analysis (PCA) training examples, the ASM represents the variations of contours and distribution of grey-level values around such contour. Fitting the ASM to a target image can thus find the desired target object. The ASM model is sensitive to starting points of localization. If the given starting point is located far away from the central point of the target, the model will be unable to locate the correct object.
A customized version of the Generated Hough Transform (GHT) based on the pose estimation of the vertebrae it an effort proposed by Tezmol and it basically developed in terms of initialization for the customized ASM, witch was originally formulated by Cootes. The GHT algorithm also requires prior information and knowledge of the shape to be segmented and it's a template matching technique, based on the invariant to the variation in the scale and rotation. An effort to customized version of Generated Hough Transform was extended to segment lumbar vertebra Gururjan. The basic GHT formulation for detection of arbitrary shapes in images was given by Ballard and the GHT of the NNESII x-ray images as been effectively investigated by Tezmol and Gururjan. The GHT algorithm consist of lookup table called R-table for an arbitrary shape, the geometry for building the
R-table consists of a reference point, which is the original of an axis system fixed in the template shape. The pairs of (r, α) are an arbitrary point on the template boundary, where r is the Euclidean distance from the reference point to the boundary point and α is the angle between the connecting line to the reference point.
Investigation Of Prototypes In CBIR
Te ultimate goal of medical image retrieval system is to deliver the similar images compared to a query image in most effective and efficient way, whether CBIR or text based and is gaining importance as support tool for diagnosis, research and education in the medical domain. The medical images characteristics such, size, resolution, multi modality, data Heterogeneity (2D, 3D), structural and functional context make processing, organization and interpreting much difficult in automated manner  and may impose challenges for efficient processing and indexing.
During the last decade, several useful research prototypes have been implemented and some are still ongoing in developing medical image databases and indexing techniques, especially content-based image retrieval (CBIR) systems. Some specific requirement for research was pointed out at National Institute of Health (NIH) workshop sponsored by the national cancer institute convened on the specific topic of medical image database towards the direction of content-based retrieval systems. These prototypes and projects differ according to their purpose and image modality, among them few important project and prototypes are ASSERT, CBIR2, IRMA, I-BROWSE, Pathfinder. Section below we will discuss briefly how each of theses systems their feature and modality for which it applied.
The ASSERT system is designed for high resolution computed tomography (HRCT) images of the lung, the system include attribute that measure the perceptual properties of anatomy, such as linear and reticular opacities, nodular opacities diffuse regions of high attenuation and diffuse regions of low attenuation. Many attributes are derived such as the thickness of the bronchi walls, the number of the cells, the average cell size and the number of the cells adjacent to the lung boundaries or fissures etc, by applying various high, low and gray scale threshold values.
CBIR2 system is a research project, aimed for retrieval of spine X-ray images. The indexing system includes methods for automated image segmentation, image feature extraction, feature vector computation, feature organization, and text data organization. In this project, twenty five biomedical features of interest in the spine x-ray images are considered, however, only three features are reliably and consistently detected, viz., anterior osteophytes, disc space narrowing, and subluxation for the cervical spine and spondylolisthesis for the lumbar spine for perceptual based feature extraction. CBIR2 system is actually an extended version of previous text based WebMIRS system with CBIR mechanism.
Pathfinder system is a multipurpose wavelet-based image retrieval system based on the nature of wavelet transform, which preserve the location information, wavelet based CBIR systems can be considered more suitable for the medical imaging domain. It has been proven effective in searching biomedical images such as digitized pathology slides.
Image Retrieval in Medical Application (IRMA) system is following a general integration approach, which is the ultimate goal of most of the systems. This approach tries to integrate CBIR and PACS to make a profound impact on diagnostic quality supporting both evidence-based medicine and case-based reasoning. The main concept of IRMA is based on a separation of the whole process in seven steps to enable intelligent image retrieval. These steps are image categorization, image registration, feature extraction and selection, indexing as multi scale representation, identification and retrieval.
IBROWSE is another project, developed jointly by City University of Hong Kong and the clinical school of the University of Cambridge, aimed at supporting intelligent retrieval and browsing of histological images obtained along the gastrointestinal tract. With the help of knowledge bases and reasoning engines, high level semantic attributes of images are obtained and textual annotation of images are automatically generated in this system.
A prototype design for content-based functional image retrieval system for dynamic positron emission tomography (PET) images presented by , their system supports efficient content based retrieval based on physiological kinetic features. They argued that existing CBIR approaches might not be optimal when applied to functional images due to the unique characteristics of these images.