Medical X Ray Image Retrieval Health And Social Care Essay

A robust, accurate and fast medical image retrieval tool is required. To seek and recover medical image from a really big databases, in order to give their utility in diagnosing of diseases, researching and developing. Now yearss, Content based image retrieval for medical image is active research Fieldss, where the images is retrieved based on their ocular content. Several surveies were conducted to better the current work ; we think that non-linear combination of multi characteristics would be a strong betterment to build a powerful medical image retrieval tool. The non additive combination of multi characteristics is done by delegating different weight for each characteristic harmonizing to their relevancy ; this attack was applied for general CBIR and it successfully better the retrieval consequences. Several attacks to find the characteristics weight, these attacks include manual weight scene, constellating, usage of relevancy feedback, statistical attacks, nervous web, familial algorithm and Fuzzy set. In this reappraisal we provide an overview of CBIR system, a current medical CBIR and a study of the recent and different characteristics burdening attacks that have been applied to non medical images in order to hold complete apprehension of the job and to look into their pertinence to medical images.


Table of Figures


In the age of digital information revolution, the importance and the value of information depends on how easy be found and the velocity and truth of recover them. Digital images specifically are exponentially increased, this due to the big figure of images produced by different Fieldss such as the scientific, medical, geographical and other more, and the development of digital image detectors which create a immense sum of non-textual information. Therefore, a sophisticated and efficient tool to hunt and recover digital images is indispensable demand. The old attacks of image retrievals are rely on the cardinal word hunt, where each images are annotated with one or more cardinal word. Then, to recover image a text- based hunt is done to happen the images with similar key words. These attacks are non genuinely sufficient, because most of the digital images that stored in databases have no or incorrect cardinal words. The recent attacks are recovering images based on their ocular content, this attack called Content Based Image Retrieval ( CBIR ) . CBIR was active research field in many applications. Medical images CBIR is one of the interesting research country due to the big figure of digital medical image produced in infirmary and wellness attention establishments. Medical CBIR are investigated in several surveies in order to develop a powerful medical image retrieve system that would be really helpful to increase the truth of diagnosing. Few medical CBIR was developed and utilize such as ASSERT, medGIFT and FSSEM.

In General, CBIR approaches represent each image as a characteristic vector and step the similarity between images with distance between their corresponding characteristic vectors harmonizing to some metric. These feature vectors are designed to encode colour, texture and form belongingss of the image. The chief challenge is to find an appropriate image representation used for retrieval since the truth of image retrieval chiefly depends on image characteristic extraction and similarity map used. More discriminated characteristics lead to better consequence. Recently, more researches concentrating on combination of multi characteristic to stand for the image, this lead to look into how these characteristic will unite. The easiest method is to unite the partial similarities in a unvarying manner. Unfortunately, this attack may non be effectual since the influence of different similarity matrices is by and large non every bit of import. Therefore, we need to analyze relevancy weights for each similarity. Several attacks of burdening strategy are present in the literature. Some of these attacks allow the user to manually put the weight harmonizing to his penchants. Others determine the characteristics ‘ weight automatically either by constellating or sorting the image in database, so weights each characteristics harmonizing to bunch rank. Some statistical attacks are besides used in order to happen the relevancy weight. Other surveies use unreal intelligence method such as Genetic algorithm to happen the optimum weight combination. The Neural Network and Fuzzy Integral are besides used. In this reappraisal we present late available attacks of image characteristics burdening. This reappraisal is done to hold a clear and complete perceptual experience of the characteristic burdening job and the current researches and surveies were done to work out it. At the terminal, it has been proven that the combination of characteristics with different weight is bettering the retrieval consequences. We are belief that using such techniques for medical CBIR would heighten the retrieval preciseness and truth which is primary aim in medical CBIR.

This reappraisal is organized as the followers: brief background about content based image retrieval and its chief constituent in subdivision 2. Section 3 provides the importance of Medical CBIR and the chief medical CBIR systems. The recent and different characteristics burdening method in CBIR is in subdivision 4. Section 5 present the decision and future work.


A brief background about content based image retrieval, its chief constituent and the common rating public presentation is provided in the undermentioned subdivisions.

Content Based Image Retrieval

Content based image retrieval ( CBIR ) refers to group of methods which involvement to seek and recover images based on their ocular content such as colour, texture and form. CBIR has been active and fast growing field. This involvement was indispensable to manage the immense addition in digital image produced by scientific, educational, medical, industrial and more other application. And to get the better of the jobs with text-based image retrieval since most of the images stored in databases have no or incorrect associated textual information [ 1 ] . In typical CBIR, the images stored in the database are represented by one or more ocular content, these ocular content where extracted as characteristic vectors. The characteristic vectors of all images are stored to organize characteristic vectors database. At retrieval procedure, the user provide the system with query image, the system converts it into its representation of characteristic vectors. The consequence is an order list of database images is show to the user, based on their similarities step between question characteristic vector and the characteristic database. Some recent retrieval systems have introduced users ‘ relevancy feedback to set the retrieval procedure in order to obtain more meaningful retrieval consequences. Figure [ 1 ] shows an overview of CBIR system.

The undermentioned subdivisions provide description of each constituent.

Ocular Feature and Feature Extraction:

The chief measure is to find the appropriate method to stand for the ocular content. The ocular content is represented most of the clip as numerical vector, and some clip it is represented as graph such as in [ 2 ] . Some other attack represent image as “ bag of ocular words ” , where image see as papers. Then features extracted from local and planetary color/texture the image are considered as ”visual words ” . Images can so be analyzed by numbering the frequence of the meaningful ”visual words ” and represented by a ”visual words ” frequence histogram [ 3 ] .

Color is one of the basic and obvious widely used ocular characteristics in image retrieval. Color characteristics are robust and stable comparison to different ocular characteristic. Furthermore, the colour characteristic computation is comparatively simple. In general, the colour characteristic of an image is represented by colour histograms. One of chief end of colour histogram is to depict the planetary colourss distribution for image. However, some bing research used local/regional colour histogram such as in [ 4 ] [ 5 ] [ 6 ] . Where the image is spliting into sub-block and histogram of each block is calculated. [ 1 ] .

Figure: Overview of CBIR system

Texture characteristic is incorporating information of ocular forms in the image surfaces and their spacial location. It is normally obtained utilizing filter method. One of the well-known attacks for depicting texture of an image is Edge Histogram Descriptor [ 4 ] . Other texture characteristic extraction methods are co-occurrence matrix [ 7 ] , ripple decomposition [ 8 ] [ 7 ] [ 9 ] , Gabor filtrating. [ 1 ] .

Shape is another of import characteristic in image retrieval, it is extracted either by placing geometrical belongings of image parts such as country, disk shape and etc. Or it extracted utilizing border, contours, lines and etc. the form characteristic normally use some sort of image cleavage in order to place the above belongingss. [ 1 ] .

Most of the recent surveies in CBIR use the MPEG-7 forms. It was developed by The Moving Picture Experts Group ( MPEG ) , The primary aim of the constitution of this criterion is to standardise the description of images and picture stored in information bases. In order to assist figure of different applications in image retrieval procedure based on their ocular content. The forms include set of colour, form and texture forms. An overview of these forms can be found in [ 10 ] . Figure [ 2 ] shows these forms.

Although many different ocular characteristics and the diverseness of ways to pull out them, but it still low degree characteristic because it can non stand for the construct or semantic of the image as the user privation, this what is known as “ semantic spread ” . There are several extended betterments to cut down this chitchat.

Figure: MPEG-7 Ocular Forms

Similarity Measure:

A The similarity step scores the similarity between the characteristic vectors of the question image and those of the characteristic database. There exists several similarity/dissimilarity steps ; the most normally used is the Euclidean distance which scores the similarity as the followers:

( 1 )

Where Q is the query image, s is the stored image, is theA ith characteristic constituent, andA nA is the dimensionality of the characteristic vector. Other similarity step is the Manhattan distance which calculated as followers:

( 2 )

Another step used in image retrieval is Histogram Intersection [ 11 ] , which calculated as

( 3 )

The interesting fact about the Histogram Intersection is its ability to compare and fit two images with different country of two histograms. [ 12 ]

The Earth Mover ‘s Distance ( EMD ) is used as similarity step in image retrieval as the followers:

( 4 )

Where P and Q are two images with m, n characteristic respectfully, pi is the correspondence characteristic ; the weight of the bunch is wi. The signatures stand foring as ] and ] . D is the land distance matrix where dij is the land distance between images pi and qj, fij is the flow between pi and qj, which minimizes the overall cost and computed as the following

( 5 )

EMD can fit on partial similarity, and it is really efficient to utilize in image retrieval [ 12 ] . In above steps a smaller distance indicates a higher grade of similarity between the two compared images.

Vector Cosine Angle Distance is besides used to hit the similarity between tow vectors the Cosine of the vector angle between first and 2nd vector in n dimension. The cosine of the angel is calculated as

( 6 )

The value of cosine is between 0 and 1, 1 agencies indistinguishable vector and 0 agencies that the vectors are different.

Relevance Feedback:

Relevance feedback constituent is introduced to recent CBIR to increase the truth of retrieval consequences and to cut down the semantic chitchat. The users ‘ feedbacks determine the grade of satisfaction of the initial consequence provided by the system in response to query image. These negative and positive feedbacks are adopted to polish the current consequence, change the characteristic weight, modify the hunt parametric quantity, or to better the consequences for following retrieval by hive awaying the feedback information [ 13 ] .

CBIR Performance Evaluation:

Many retrieval public presentation steps are used by several surveies and researches. In this subdivision, a brief description of presently used methods to mensurate public presentation the retrieval procedure provided. The widely used rating steps are in general information retrieval system are preciseness and callback. The precisionA is the fraction of retrieved cases that are relevant which measure the ability of the system to recover all relevant cases. RecallA is the fraction of relevant cases that are retrieved which measure the ability of the system to recover merely relevant cases. Both steps are calculated as the following

The preciseness and remember normally represented as Precision vs Recall graph which is easy to disrupt. Some of retrieval system evaluated utilizing Average Precision merely or remember but it is difficult to disrupt. The callback and preciseness step does non care about the order of recovering. Another information retrieval rating steps include mistake rate which is the fraction of non-relevant cases that are retrieved.

The Average Normalized Modified Retrieval Rank ( ANMRR ) A is retrieval public presentation step which considers the order of retrieved consequence. The computation of ( ANMRR ) A as the followers: For each question a set of land truth images that are most relevant to the question were identified. Let the figure of land truth images for a given question Q be NG ( Q ) . First K retrievals ( the top ranked K retrievals ) are examined, where

K=min ( 4*NG ( Q ) , 2*max ( NG ( chi ) ) .

For each land truth image k that was retrieved in the top K retrieval, a rank value Rank ( K ) is attached. The Rank ( K ) is the retrieval rank of the land truth image is

Where R ( K ) is the rank of image K in retrieval consequences. The mean rank AVR ( Q ) for question Q is computed as follows:

The modified retrieval rank as followers:

Normalized MRR is

The NMRR ( Q ) has values between 0 agencies perfect retrieval and 1 agencies nil found. Finally, for the whole question sets,

ANMRR is defined as follows:

Content Based Medical Image Retrieval

Medical images are of import information beginning in many Fieldss in wellness attention such as diagnosing, surveies, research and acquisition. Nowadays, medical images are in digital signifier, 1000s figure of digital medical images are produced every twenty-four hours, and computing machine systems to hive away and recover images such as Picture Archiving and Communication Systems ( PACS ) are available. Therefore, unsnarling utile information from these images needs an efficient tool for image retrieval. The CBIR for medical images grab more attending and became one of chief and basic application of CBIR. Several surveies and system have been proposed, nevertheless, merely a few content-based retrieval systems have been developed such as ASSERT. In the undermentioned subdivisions we describe the importance of CBIR in Medical image, the chief medical CBIR systems and the chief Visual features that extracted from medical image.

Importance of Medical CBIR:

Several surveies define the end of the medical information system as “ deliver the needed information at the right clip, the right topographic point to the right individuals in order to better the quality and efficiency of attention procedures ” . For medical image which is of import portion of medical information, such a end will more likely to carry through if there are a strong and efficient tool for image retrieval. The 2nd ground is the outgrowth of the usage of Decision Support Systems in radiology and computer-aided nosologies for radiological. Besides create a demand for powerful image retrieval tool. The immense size of stored medical image and the high mistake rate of text associated with medical image heading require incorporating and utilizing the techniques of CBIR in medical image hunt and retrieval in order to hold the needed consequences with high preciseness and truth. The usage of CBIR in Medical images to increase the diagnostic quality was proven. [ 14 ] [ 15 ] . In add-on, one of the most of import countries which expected to better is instruction and research in medical images. That due to the hunt based on the ocular characteristic of image can happen and recover related images with same or different diagnosing [ 15 ] .

Medical CBIR systems:

Content based image retrieval for medical images have been research topic in recent old ages. Several survey and research have been proposed. However, merely few Numberss of systems has been developed and used. This subdivision briefly describes the major medical CBIR system.

Automatic Search and Selection Engine with Retrieval Tools ( ASSERT ) :

This system is developed to seek and recover images from High- Resolution Computed Tomography ( HRCT ) database specifically image of lungs. The system use physician-in-the-loop attack which requires the user engagement to find the pathology-bearing parts and place certain anatomical landmarks for each image. ASSERT used about 255 characteristics texture, form extracted from general lungs characteristic and assorted perceptual classs for each pathology bearing parts ( PBR ) . Figure [ 3 ] show the chief GUI of the ASSERT system [ 16 ] . The ASSERT undertaking has besides performed a survey on the usage of content-based image retrieval as a diagnostic assistance. And it showed an betterment in diagnostic quality, particularly for less experient radiotherapists [ 14 ] .

Figure: Main GUI of ASSERT


The system uses the unfastened beginning image retrieval system ( GIFT GNU Image Finding Tool ) for the retrieval of medical images in the medical instance database system. The medGIFT retrieval system extracts planetary and regional colour and texture characteristics, including 166 colourss in the HSV colour infinite, and Gabor filter responses in four waies each at three different graduated tables ; Combinations of textual labels and ocular characteristics are used for medical image retrieval. Figure [ 4 ] show the chief interface of medGIFT system. [ 14 ]

Figure: medGIFT web interface

NHANES II ( The Second National Health And Nutrition Examination Survey ) :

NHANES II is a system implemented to seek and recover images of Spine X-ray. The form of vertebra is obtained by Active Contour Segmentation ( ACS ) tool, which allows the users to make a templet by taging points around the vertebra. The templet is used to come close the polygon form of vertebra. Then, the approximated curve of vertebra is converted to tangent infinite for similarity measuring [ 17 ] .

FSSEM ( characteristic subset choice utilizing expectation-maximization bunch )

FSSEM is a content-based image retrieval use a new hierarchal attack called “ customized-queries ” , the system usage multi characteristic vectors alternatively of one characteristic vectors. The retrieval procedure is done in tow stairss, foremost the question image is sorting based on subset of most discriminatory characteristics vectors. The 2nd measure is to recover the list of most similar images from the predicted category. [ 17 ]

Medical X-ray Image Visual Features:

As mentioned above the ocular characteristic choice and extraction is the key to build high public presentation CBIR. Therefore, many surveies and research proposed to accomplish a good CBIR system for medical images. The procedure of extract ocular characteristic of medical image is disputing. Medical image have different mode. Each of this mode has its ain ocular characteristic which different from others. Besides medical images are gray degree which means that the colour characteristic is non really efficient. Most medical image has dark background contain no information and bright foreground. In add-on the most discriminatory characteristics are the 1 who extracted at local/regional degree. [ 18 ]

In this subdivision we provide some surveies that consider CBIR for medical X-RAY images. Some of the X-ray CBIR is utilizing general ocular content such as Color, Texture and Shape for illustration the For illustration, Shim et al [ 19 ] proposed an algorithm for X-ray image categorization and retrieval utilizing MSVM with an ensemble characteristic vector by uniting a colour construction form ( CSD ) based on the Harris corner sensor to pull out colour fluctuation in foreground merely. For texture characteristic an border histogram form of the image of MPEG-7. Bhattacharya et Al In [ 20 ] extract the characteristic vectors utilizing a colour bed form and a histogram form of MPEG-7 standard forms. An efficient content based medical image retrieval strategy is proposed in [ 21 ] , an automatic classification method for X-ray images, utilizing the Gaussian mixture mold ( GMM ) and Kullback-Leibler ( KL ) step, the purpose of which is to pre-label image classs harmonizing to organic structure parts. To pattern the image class, set of strength and texture characteristic are extracted, and so harmonizing to these characteristics the pels are grouped to organize a different part by GMM and each part represented as blobs. To gauge the matching tonss between images KL step is used. Figure [ 5 ] show X ray of manus and GMM blobs.

Figure: GMM representation of manus [ 20 ]

Feature Weight:

Normally, In CBIR uniting multi characteristics is indispensable demand. Since one characteristic is non know aparting the images content. In several surveies it has been proven that good combination of characteristics would better the public presentation of image retrieval consequences. In such attack, the entire similarity of relevant images is computed by fused the similarity of multiple characteristics in one planetary matrix as the followers:

( 7 )

Where vitamin D ( q, J ) is the entire similarity between the question image and the stored image, di characteristic distance between the question image and an image in the database. The hunt job here is to happen the influence of different characteristic on the retrieval consequences. The traditional attack is to utilize one planetary matrix that combines the partial similarities in a unvarying manner, presuming the influence of each characteristic is equal. Unfortunately, this attack may non be effectual since the influence of different similarity matrices is by and large non every bit of import in the definition of the class to which similar forms belong. Therefore, happening the appropriate relevancy weights for each characteristic is necessary in order to increase the public presentation of retrieval consequences. Therefore the planetary distance matrix is calculated as the followers:

( 8 )

Figure [ 6 ] show features combination in CBIR.

Figure: Feature Combination

Several attacks of characteristic weight computation are proposed for general CBIR. The simplest attack is to put the relevancy characteristic weight manually by the user. In [ 2 ] the proposed system gives the user the ability to burden each property harmonizing to its relevancy. Where the image is represented by a Fuzzy Attributed Relational Graph ( FARG ) , FARG describes each object in the image, its properties and spacial relation. Six nodes properties where used, which are Label, Size, Coarseness, Contrast, Directionality and Color. Other manual characteristic weight is proposed in [ 22 ] , the proposed method introduce the usage of fuzzed logic in image retrieval, the image is represented by colour histogram as colour characteristic and the minute invariants is extract as form characteristic. Similarity step between question image and database image based on distance of colour and form vector are converted into three fuzzed variables including “ really similar ” , “ similar ” and “ non similar ” . After mensurating the characteristics difference in term of fuzzy variable, a regulation based of nine regulations is obtained. And the user is weight the regulations harmonizing to his penchant. The trade market experts in [ 23 ] find the weight vector of different constituent of similarity step. Another attack is to by experimentation prove the public presentation of different combination until reach the best retrieval consequences such as in [ 8 ] . The consequences show that the combination of leaden characteristic has higher preciseness compared to merely combination of characteristic without weights. Despite the betterment in image retrieval consequences but it will go a hard procedure to the user when the figure of characteristics extracted is increasing, increasing the figure of extracted characteristics requires finding of its weight by the user. Therefore, automatic weight finding is required.

Number of different methods has been investigated to automatize the procedure of characteristic burdening. One of these methods is image constellating. Frigui et al [ 11 ] proposed CARD, a fuzzed bunch algorithm that is able to partition objects taking into history multiple unsimilarity matrices and that learns a relevancy weight for each unsimilarity matrix in each bunch. The CARD algorithm has two versions which are extensions of the RFCM and FANNY constellating algorithm. The writers mention one of import usage of the CARD in CBIR because it automated the computation of characteristic weight for each bunchs and non necessitate user intercession. In their attack the weight and the ratio of the scattering of specific distance matrix to the entire scattering are reciprocally related. This mean more compact bunch in specific distance matrix will give it a higher weight. The weight of each characteristic is a measure in the bunch algorithm and it is calculated as the followers:

( 9 )

( 10 )

Where stand for the grade of rank to the bunchs is the distance between J and K points. The steps the scattering of the ith bunch harmonizing to the distance matrix Rf. The steps the entire scattering of the ith bunch sing all distances matrix.

Another method of automatic characteristic finding is based on iterative weight update. This method proposed in [ 3 ] , the proposed system extract set of characteristics from labelled preparation images. Each characteristic is assigned an initialized weight 1, which will be iteratively updated to choose the discriminatory characteristics. The updated procedure is done after the similarities between images are computed utilizing Earth Mover Distance. The N most similar images are selected as the initial searching consequences. Then, use these initial searching consequences to update the characteristic weights. The weight procedure is done by calculating characteristic fiting part matrix which measures the importance of peculiar characteristic degree Fahrenheit to image ; the part matrix is calculated as the followers:

( 11 )

Where the dij, fij is the flow and land distance which computed in EMD and is normalising factor and computed as: . The new weight for is the mean for all related parts

( 12 )

Then, the resulted discriminatory characteristics are used to build the ”visual words ” . Finally, bag-of-feature method is used for refined image retrieval.

Another automatic characteristic burdening which proposed by Wilkins et al [ 4 ] . The proposed method finding the characteristic weights automatically at query clip. The proposed attack computes the weight of different characteristic by calculated the ratio of distance between consequences steps in a top subset of the consequences. The chief premise of this method is:

“ That given a set of normalized similarity tonss, if we calculate the mean difference between next tonss across the consequence set, and so compare that to the mean distance among the really top subset of the ranking, we would anticipate to happen that the mean of the top subset shows a tighter bunch of similarity tonss than that of the larger set. But that the closer these values, the more likely that there is similar tonss at a greater deepness. ”

The average mean distance of the consequences set with size UB calculated as:

( 13 )

The similarity bunch ratio between MAD value of peculiar subset and the one of the larger set ; represents the grade of constellating which have similar or same similarity mark and it is defined as:

( 14 )

To cipher the characteristic weight, merely cipher the mark for each characteristic and find the comparative per centum of that mark, against the amount of the tonss.

( 15 )

The proposed method was test on two different characteristics, one for colour and the other for texture. And the consequences show little betterment in retrieval consequences in term of retrieval preciseness and callback. One of chief advantage of this method it non necessitate preparation or user intercession.

Relevance feedback is besides used in automatic characteristic weighting. The writers of [ 24 ] proposed a new theoretical account which integrates keyword-based and content-based searchesA in off where the relevancy feedback consequence about all images that have the similar characteristic vector. At retrieval stage system receives feedback, analyze the RF images and return rearrange list of consequences. RF is besides used to modify the keyword assurance of the images, raising the assurance degree of positive images and take downing that of negative images. The assurance degree alteration is applied to all images with the similar characteristic vector. Another feedback based attack proposed in [ 25 ] , the weights of the characteristics are calculated based on the positive and negative labeled images by the user. If the positive images have common form which belong to the same bunch, so the characteristics have higher weight and the same done to negative images. A Nervous Network is used in [ 9 ] . The nervous web was trained utilizing a set of similar and non similar images. Once the web is trained, the characteristic categories have the proper weights so that they can be used in uniting heterogenous characteristics. The user feedbacks are used to set the weights. Using relevancy feedback in burdening will guarantee the user satisfaction of retrieval consequences.

Familial algorithm is besides used in order to happen the appropriate characteristic ‘s weight combination [ 7 ] , Shao et Al, change the finding of characteristic weight to optimisation job and usage GA to happen the best weight combination in order to acquire accurate retrieval consequences. In [ 5 ] the Fuzzy Integral is used for the weight value of each characteristic, in comparing of fuzzy-integral similarity, fuzzed step of each set of power set contain elements represent each characteristic is computed by experimentation. The weight values for each characteristic and combination of characteristics are computed based on fuzzed step. The concluding similarity, consequences from what generation the original similarity by deliberate weight values.

All above discoursing methods are tested and evaluated on general image except for [ 7 ] which tested on CT images. For Medical X-ray image [ 20 ] the characteristic weight are specified by the user.

The following tabular array summarizes the characteristic burdening attacks for CBIR. It shows characteristics representation, type of images tested on, the user intercession, the usage of RF in the weighting, if the method required preparation stage to larn the weight or non and what to weight.

Table: Feature burdening method for CBIR

Method in

Feature Representation



User intercession

Relevance feedback


What to burden

[ 2 ]






[ 22 ]





fuzzy regulation

[ 23 ]





distance matrix

[ 11 ]





distance matrix

[ 3 ]





distance matrix

[ 4 ]




distance matrix

[ 24 ]



Relevance feedback



keyword assurance

[ 25 ]



Relevance feedback


distance matrix

[ 9 ]



Nervous web



distance matrix

[ 7 ]


Nutmeg State images


distance matrix

[ 5 ]



Fuzzy built-in

fuzzed step

[ 20 ]






class/cluster rank

Decision and Future work

In this work, we were involvement to analyze the relevancy weight of multi characteristic for medical CBIR. However, few Numberss of proposed surveies were applied to medical image. Therefore, we study the current work done and applied to non-medical in order to compare, survey, and happen the current work done in this field. For future we intend to compare between these methods and look into their pertinence to medical CBIR.