Skip to content
BY-NC-ND 3.0 license Open Access Published by De Gruyter December 9, 2016

An Efficient Multiclass Medical Image CBIR System Based on Classification and Clustering

Mahabaleshwar S. Kabbur

Abstract

In this paper, we are going to present the multiclass medical image content-based image retrieval (CBIR) system based on classification and clustering. Images are segmented using hill climbing-based segmentation (HCBS) based on the extracted visual features. In the improved HCBS technique, a clustering that is based on kernel-based fuzzy C-means is employed. In the next step, features like color, texture, edge density, region area, and visual words from the segmented images are extracted. The visual word can be extracted by using the clustering techniques. This visual word represents the uniqueness of the medical image, and it is used for better classification. Then, the image can be classified by using an optimal classifier artificial neural network based on the firefly algorithm. This classification leads to filtering out the irrelevant images from the database and reduces the search space for further retrieval process. In the second stage, the relevant images are extracted from the reduced database based on the similarity measure. The proposed CBIR technique is assessed by querying different images, and the retrieval efficiency is estimated by determining the precision-recall values for the retrieval results.

1 Introduction

Content-based image retrieval (CBIR) applies to techniques for retrieving similar images from image databases, based on automated feature extraction methods. In recent years, the medical imaging field has been growing and is generating a lot more interest in methods and tools, to control the analysis of medical images. To support clinical decision making, many imaging modalities, such as magnetic resonance imaging (MRI), X-ray computed tomography (CT), digital radiography, and ultrasound, are currently available [9]. CBIR also has been applied to biomedical image retrieval [2].

Clinicians and medical researchers routinely use online databases such as MEDLINE to search for bibliographic citations that are relevant to a clinical situation. The biomedical evidence they seek is available through clinical decision support (CDS) systems (CDSSs) that use text-based retrieval enhanced with biomedical concepts. Authors of biomedical publications frequently use images to illustrate medical concepts or to highlight special cases. These images often convey essential information and can be very valuable for improved CDS and education. Text-based retrieval of images has, thus far, been limited mostly to caption and/or citation information. To be of greater value, images in scientific publications need to be first annotated (preferably automatically) with respect to their usefulness for CDS, to help determine relevance to a clinical query or to queries for special cases important in educational settings [4, 5, 12].

This article discusses a method for multimodal image annotation that utilizes (i) image analysis techniques for localization and recognition of author-provided overlays on the images; (ii) image feature extraction methods for CBIR; (iii) natural language processing techniques for identifying key terms in the title, abstract, figure caption, and figure citation (mention) in the article; an (iv) use of structured vocabularies, such as the National Library of Medicine’s Unified Medical Language System, for identifying the biomedical concepts in the text [12].

As discussed in earlier works [1], these steps can be used to associate the biomedical concepts in the text to specific regions in the image. The relevance to a clinical query is aided by this addition of semantic information to extracted image features for improved CBIR. Traditionally, CBIR tends to be limited to use of visual features in identifying similarity among a collection of images. This has spurred discussion on the “semantic gap” [6] that is introduced when high-level concepts are represented through low-level visual features such as image color and texture (for example). Such a semantic gap can be minimized through annotation by biomedical concepts that are extracted from the article text and applied to relevant regions within an image.

General CBIR also could be improved by the proposed approach in a similar manner as text-based retrieval is improved. In this case, no text information is available but only visual features are used. The CBIR identifies relevant articles as text-based retrieval does in the multimodal method. Annotations and regions of interest (ROIs) in retrieved images can be identified by the annotation recognizer and then be used to re-rank the results [12].

At present, images needed for instructional purposes or CDS appear in specialized databases or in biomedical publications and are not meaningfully retrievable using primarily text-based retrieval systems. Our goal is to automatically annotate images extracted from scientific publications with respect to their usefulness for CDS. A future CDSS could then provide images relevant to a clinical query or to queries for special cases important in educational settings. An important step toward attaining the goal is automatically annotating images and related text [11].

2 Literature Review

A lot of works have been proposed by researchers for biomedical article retrieval. A brief review of some of the recent researches is presented here.

McDonald et al. [8] proposed a classifier that was designed specifically to be readily adaptable to a wide domain of knowledge. For the identification of articles potentially mentioning genomic variations or mutations of a specific gene, the system requires only (i) the classifier; (ii) a set of training articles or abstracts that contain both positive and negative instances of the type of genomic mention of interest; and (iii) genomic variation tagger. Preliminary results have shown that performance is slightly but not substantially improved with the addition of the tagger. Furthermore, the classifier can be trained upon any set of documents in which a contextual distinction can be made, although the performance will likely vary depending on how precisely the distinction between positive and negative instances can be defined.

Torii and Liu [10] proposed a simple and easy-to-deploy classifier ensemble approach for biomedical document classification/retrieval tasks. In the proposed approach, constituent classifiers were built by varying the sizes of the feature set for a machine learning algorithm. Note that even when a single classifier is employed in a database curation project, a number of classifiers with different sizes of feature sets would be built anyway before the best performing system is selected. The proposed approach suggests combining such intermediate classifiers. In our experiments, support vector machine (SVM) ensembles outperformed all the constituent classifiers in terms of both AUC and BEP. Using this approach, we updated the classification performance previously reported on the benchmarking data sets, and set new baseline performance of the data sets. However, the ensemble approach was not effective when there were no sufficient data to train reliable constituent classifiers or when it was applied to naïve Bayes classifiers.

Cheng et al. [3] proposed a model in which figure image segmentation is an important and necessary first step in annotating images for improved information retrieval for CDS. This step helps subsequent image annotation and CBIR methods to be performed optimally. For accurate subfigure image segmentation, it is first needed to detect the image type. Regular images usually provide a strong interpanel boundary that is used to detect the subfigure panels. Finding subfigure panels in illustration images is more challenging.

You et al. [11] proposed a model in which detecting arrows, pointers, and other annotations such as text labels can be very beneficial in locating ROIs within figures in biomedical articles. Such annotations can be identified through relevant text snippet analysis (captions, figure mentions in the article text). Image analysis methods are necessary to identify the location of the symbols in the figure images. Identifying these and the image content annotated can be valuable for improved biomedical retrieval. This is achieved by attaching biomedical concepts extracted from caption and mention text analysis to image features computed at the image ROIs annotated by these symbols. Such tagging can help improve image indexing quality and subsequently the indexing and retrieval of biomedical articles through both text-based retrieval methods as well as CBIR.

You et al. [13] proposed a model in which authors frequently use pointers and symbols to highlight specific local regions and mention them in figure captions and text discussions. Localizing those pointers can help extract specific local ROIs, and using the ROIs could improve the relevance quality of conventional retrieval approaches by combining textual and image features from local regions. Research was conducted to enhance our prior work on pointer recognition and ROI extraction. A region growing technique was applied to improve pointer segmentation and ROI extraction performance. An active shape model-based pointer recognizer was developed to handle pointers that cannot be recognized by the Markov random field recognizer due to some distortion in their boundary. An average 87% success rate on pointer recognition was achieved.

From the above literature survey, we conclude that for effective retrieval of image, we focus on the classifier. We have to choose the classifier in such a manner that will classify the images in an appropriate and efficient way. Secondly, for accurate annotation of images, we also have to focus on its type. ROI is another best way that is valuable for improved biomedical retrieval. This will reduce the size of the database. Thus, it will take less time for image retrieval.

3 Problem Definition

CBIR has been one of the vivid research areas in the field of computer vision over the past decade. In the medical field, images, especially digital images, are produced in ever-increasing quantities and used for diagnostics and therapy. Content-based access to medical images for supporting clinical decision making has been proposed that would ease the management of clinical data, and scenarios for the integration of content-based access methods into Picture Archiving and Communication Systems have been created. CBIR extracts the images that are suitable to the specified query image from large image databases based on the image features. However, retrieving images based on the features extracted from the large database is a tedious task. Even though the task has been accomplished, extracting a limited number of features leads to ineffective retrieval. The majority of the existing methods are developed around a specific imaging modality, and retrieval methods in these systems are task specific. Only a few systems have a goal of creating CBIR systems for heterogeneous image modalities. To enable effective search in a large medical image collection of diverse modalities, it might be advantageous for a retrieval system to be able to recognize the image class prior to any kind of post-processing or similarity matching. In order to overcome these issues, we propose an efficient CBIR system based on classification of image modalities, which will reduce the search space to enhance the efficiency. For effectual retrieval of images, a CBIR system with a dual-stage retrieval process is proposed.

  • The main purpose of feature selection is to reduce the number of features used in classification while maintaining acceptable classification accuracy.

  • In feature selection, every feature is observed. The high number of features is a great obstacle for classification. Therefore, a feature dimension reduction method is applied to reduce the feature space without losing the accuracy of classification.

  • We decrease the number of features and take away the unrelated, redundant, or noisy information.

  • Moreover, this develops the performance of information classification with speeding up of the processing algorithm. In our work, we develop the modified neural network classification.

  • We include a shape feature by improving the performance and efficiency of our proposed study.

3.1 Proposed Methodology

3.1.1 Hill Climbing-Based Segmentation

In the first stage, the image will be categorized based on the visual features extracted. For that, every database image is segmented into different regions using hill climbing-based segmentation (HCBS). The HCBS is a simple and fast nonparametric algorithm that is used to detect the peaks of clusters in the global three-dimensional color histogram of an image. The algorithm associates the pixels of a detected cluster based on the local structure of the cluster. The proposed algorithm efficiently detects the peaks of cluster in the global three-dimensional color histogram of an image. The hill climbing algorithm is described below (Figure 1).

Figure 1: Hill Climbing-Based Segmentation.

Figure 1:

Hill Climbing-Based Segmentation.

In our proposed algorithm, the input will be an image. After applying the proposed algorithm, the output will be a set of visually coherent segments.

  1. Calculate the color histogram of an image.

  2. Start with a non-zero bin of the color histogram and continue the uphill moves until reaching a peak as follows:

    1. Compare the number of pixels of the current histogram bin with the number of pixels of the neighboring (left and right) bins.

    2. The algorithm makes an uphill move toward the neighboring bin with a larger number of pixels, if the neighboring bins have different numbers of pixels.

    3. If the instant neighboring bins have the same numbers of pixels, then the algorithm checks the next neighboring bins, and so on, until two neighboring bins with different numbers of pixels are found. Then, an uphill move is done toward the bin with the larger number of pixels.

    4. The uphill climbing is repeated until accessing a bin from where there is no possible uphill movement. It is the stage when the neighboring bins have smaller numbers of pixels than the current bin. Hereof, the current bin is known as a peak [7].

  3. Select another unclimbed bin as a starting bin and perform step 2 to find another peak. This step is continued until all non-zero bins of the color histogram are climbed (associated with a peak).

  4. The known peaks depict the initial number of clusters of the input image; thus, the number and values of these peaks are saved.

After applying the above algorithm, an image will be segmented into different visually coherent segments. The next step is clustering of different visually coherent segment parts.

3.1.2 K-means Clustering

The visual word can be extracted by using the clustering techniques. This visual word represents the uniqueness of the medical image, and it is used for better classification. K-means clustering is the efficient algorithm for clustering the segmented image. In proposing an approach, firstly provide the rule on which the cluster is made. In the second step, the clusters are obtained based on some similarity. K-means is the simplest algorithm to solve the clustering problem. K-means take the input of the segmented image and cluster it into different groups based on a defined procedure. For achieving efficient clusters, we have to take random centers on which different clusters are made. These centers should be placed in a devious way because different positions cause different results. Thus, for obtaining the better result, it is necessary to place these centers as far away as possible from each other. We have to place those points in the same cluster, which will be near the respective centroid taken. When no point is incomplete, the first step is finished and an early group age is done. At each point, we have to re-calculate k new centroids as the barycenter of the clusters resulting from the previous step. After calculating these k new centroids, we have to calculate a new binding between the same data set points and the nearest new center. A loop will be generated. The result of these will be that the k centers change their location step by step until no more changes occur. The main aim of the given algorithm is to minimize an objective function known as a squared error function given by the following formula:

(1)J(V)=i=1cj=1ci(||xivj||)2,

where ||xivj|| is the chosen distance between xi and vj; ci is the number of data points in the ith cluster; and c is the number of cluster centers.

K-means clustering is an unsupervised learning method of cluster analysis that aims to partition n datasets into k clusters in which each dataset belongs to the cluster with the nearest mean. One thing that has to be remembered is that centers will be chosen randomly and updated each time. Here, we present the procedure of K-means for obtaining an efficient clustering.

Let X={x1, x2, x3, …, xn} be the set of data points and c={c1, c2, …, cm} be the set of centers.

  1. First, we have to randomly select c cluster centers.

  2. Estimate the distance between each data point and cluster centers.

  3. Define the data point to the cluster center whose distance from the cluster center is minimum of all the cluster centers.

  4. Recalculate the new cluster center using the following formula:

    (2)vi=(1ci)j=1cixi,

    where ci represents the number of data points in the ith cluster.

  5. Recalculate the distance between each data point and new cluster centers.

  6. If no data point was redefined, then stop; otherwise, repeat from step 3.

3.1.2.1 Feature Extraction

After segmentation, the images are split into regions. We have to extract the features from each region in the image. These features are used for classification of image modalities. Here, we will extract the correlation, contrast, multi-texton, energy, and homogeneity features.

  1. Correlation:f1=ij(ij)p(i,j)αxαyσxσy,

    where p(i, j) is the (i,j)th normalized gray tone spatial dependence matrix, and αx, αy, σx, and σy are the means and standard deviations of px and py.

  2. Contrast:f2=n=0Na1n2(i=1Naj=1Nap(i,j)),

    where Na is the number of distinct gray level in quantize image.

  3. Energy:f3=ijp2(i,j).

  4. Homogeneity:f4=ijp(i,j)1+|ij|.

  5. Multi-texton histogram

    We can easily find the difference between texture images globally. Texton is nothing but a set of emergent pattern or blobs that share common information all over the image. We emphasize on critical distances (D) between texture elements on which the computation of texton gradients depends. If the adjacent elements lie within the D-neighborhood, then only textures are formed. For clearly understanding texton detection, we are using a 2×2 grid. The 2×2 grid will represent the four pixels a, b, c, d, which are shown in Figure 2.

Figure 2: Four Texton Types Defined in MTH.(A) 2×2 Grids; (B) T1; (C) T2; (D) T3; (E) T4.

Figure 2:

Four Texton Types Defined in MTH.

(A) 2×2 Grids; (B) T1; (C) T2; (D) T3; (E) T4.

These four texton types are denoted using the symbols T1, T2, T3, and T4, respectively. For the color image C(x, y), we employ the 2×2 block from left-to-right and top-to-bottom everywhere in the image to detect the texton with two pixels as the step length. If a texton is identified, the original pixel values in the 2×2 grids are kept unaltered. With the contrary, it will have zero (0) value. At last, we obtain a texton image, denoted by I(x, y).

On performing K-means clustering, we obtain different clustered parts. These different clusters need to be classified. For this reason, we are using artificial neural network (ANN). ANN will classify the above clustered parts into some valuable classification. One other advantage of using ANN is that it also validates the different.

3.1.3 Firefly-Neural Network Optimization

3.1.3.1 Firefly

In this optimization phase, the ontology alignment by ANN classified emotions are optimized with the aid of firefly-neural network, as briefly explained below.

  • Solution representation

    The solution demonstrates with the firefly algorithm accomplishment. In the population, one firefly (solution) is a probable solution. The initial population of fireflies is created randomly for the firefly algorithm. The initial population of size Y is defined as follows:

    (3)Y=Ad  (d=1,2,...,n),

    where n is the number of fireflies.

    The initialized continuous position values are created by using the following equation:

    (4)uk*=u+min(umaxumin)*r,

    where xmin=0, xmax=1, and r represent a uniform arbitrary number between 0 and 1.

  • Fitness evaluation

    The fitness function is represented in the corresponding present investigation. The optimization formula is indicated in Eq. (5), which depends on the minimization of the objective function as given below:

    (5)W(y)=mini=1mw(yi)Hx(yi),

    where Hx(yi) is the entropy and w(yi) is the weight of the entropy of each attribute.

  • Firefly updating

    The movement of the firefly p, when attracted to another more attractive (brighter) firefly q, is evaluated by using Eq. (6), given below:

    (6)up=up+γ(r)(upuq)+ϕ(rand12).

    The second term in Eq. (7) is on account of attraction; the third term introduces randomization with ϕ being the randomization parameter; and rand is a random number produced evenly disseminated between 0 and 1.

    (7)Attractiveness: γ(r)=γ0eθrm,  m1,

    where r represents the distance between two fireflies; γ0 denotes the initial attractiveness of the firefly; and θ reveals the absorption coefficient.

    (8)Distance: rpq=upuq=k=1d(up,suq,s)2,

    where up, s represents the sth component of the spatial coordinate of the pth firefly and d denotes the total number of dimensions. Also, q{1,2,...,Fn} characterizes the arbitrarily selected index. Although q is evaluated arbitrarily, it must be different from p. Here, Fn corresponds to the number of fireflies. The hybrid of firefly and neural network is utilized to optimize an emotion speech signal.

3.1.4 Multiclass ANN for Classification of an Image

The clustered image is classified using ANN. The output of K-means clustering algorithm is taken as the input for ANN. An ANN is a feed-forward back-propagation neural network. The classification results should provide some feedback to the data set. If the results of a classifier reflect the nature of the data set better, it would be potentially more useful and powerful. The neural networks are flexible models that allow treating the multiclass difficulties in a direct way; however, the SVM models cannot allow this task in a straight way, as they are two-class models. Here, we are using multiclass ANN because it will identify the valid type of input as well as the type of respected valid input. An ANN is an oriented and effected graph. The nodes of this graph are simple automatons named formal neurons, which possess an internal state that represents their activation and by which they influence the other neurons of the network.

The case of the whole network is natural of its neuron’s activation, and by the matrixes of the synaptic weights joining a layer to the following, each one is a matrix in which we observe the weights of the links. The answer pi of a neuron number i is given as follows:

(9)Pi=f(neti)=f(j=1nxjwji),
(10)with f(x)=11+ex.

ANN consists of three layers, which are the input layer, hidden layer, and output layer. A simple structure of multiclass ANN as outlined in Figure 3.

Figure 3: Neural Network Architecture.

Figure 3:

Neural Network Architecture.

Figure 3 shows the structure of ANN. In the given ANN, inputs are taken as Ip and Is. An ANN consists of three layers. They are the input layer, hidden layer, and output layer. In the given network, we show them in the form of I, H, and O, respectively. Here, we are taking two input nodes: N hidden nodes and M output nodes. The weights of each edge are shown in Figure 2. Inputs are given at the input layer. The inputs of each node are given to each node of the hidden layer. The hidden layer is used for decision taking purpose or analysis purpose. According to the hidden layer decision, the output is obtained. In an ANN, the output layer contains more than one output. It means an output will be of a variety of types.

If the output is not what we want, then we use the back-propagation algorithm to obtain the desired output. The back-propagation algorithm uses the technique of gradient descent to minimize the distance between the wanted output and the output obtained by the network:

(11)Weights: qij=μFaqij,

where μ is called the learning rate.

The updated weights are given by

(12)qijp+1=qijp+qij.

The back-propagation algorithm can be outlined below:

  1. Initialize the random weights and initialize number of iterations n to 0.

  2. Present firstly the input vectors from the training dataset in the network.

  3. Dispatch the input vector p through the network to get an output:

    Apk=f(jqjkf(mqm)f(f(iqijxi))).

  4. Calculate the signal of errors between the real output and the desired output. Calculate the sum of squared errors and increment n.

  5. Send the error signal backwards through the network.

  6. Correct the weights to minimize the error by the following update:

    qijn+1=μδiAi+αqijn,

    where α is a constant called momentum, which serves to improve the learning process.

  7. Repeat the steps 2–6 with the next input vector until the error becomes sufficiently small or until a maximal number of iterations fixed in advance is reached.

After the classification of the images, we have to retrieve the images. Image retrieval is done by using Euclidean distance. It is also known as the L2 distance. Suppose there are two points x=(x1, y1) and y=(x2, y2). The Euclidean distance between these two points can be calculated by the formula given below:

ED=(x1x2)2+(y1y2)2.

An image is represented by a feature vector. These feature vectors are similitude using various distance measures. The lowest distances in images are ranked highest in the retrieval process.

4 Results and Discussion

The proposed CBIR system has been implemented in the working platform of MATLAB (version 7.12). The proposed system has been assessed with diverse query images, and appropriate images are recovered from the misc database. The misc database contains 10,000 images stored in JPEG format.

In our proposed methodology, we use the HCBS approach for segmenting the medical image. The HCBS algorithm efficiently detects the peaks of cluster in the global three-dimensional color histogram of an image. The output of improved HCBS is taken as the input for K-means clustering. K-means clustering is the efficient algorithm for clustering the segmented image. K-means take the input of the segmented image and cluster it into different groups based on a defined procedure. Now, we extract the correlation, contrast, multi-texton, energy, and homogeneity features using appropriate formulas. The classification of the given image is carried out by ANN. The output of the K-means clustering algorithm is taken as the input for optimal ANN. An optimal ANN is a back-propagation neural network. After applying the above steps, we retrieve the image in the appropriate format.

Figure 4 shows the image for the human lung. Part (A) shows the query image of the human lung. On applying the proposed algorithm, we get part (B), i.e. retrieval images.

Figure 4: Sample Output Obtained in the Retrieval Process for Lung Image.(A) Query image; (B) retrieved images.

Figure 4:

Sample Output Obtained in the Retrieval Process for Lung Image.

(A) Query image; (B) retrieved images.

Figure 5 shows the images of the kidney in the human body. Part (A) shows the query image of the kidney in the human body. On applying the proposed algorithm, we get part (B), i.e. retrieval images.

Figure 5: Sample Output Obtained in the Retrieval Process for Kidney Image.(A) Query image; (B) retrieved images.

Figure 5:

Sample Output Obtained in the Retrieval Process for Kidney Image.

(A) Query image; (B) retrieved images.

Figures 6A and B show the screenshot taken using MATLAB software. It shows how we take a query image to obtain retrieval images.

Figure 6: Screenshot for the Entire Proposed Methodology.

Figure 6:

Screenshot for the Entire Proposed Methodology.

4.1 Performance Analysis

The performance of the proposed CBIR system is evaluated based on precision, recall, and F-measure values.

4.2 Precision

The precision of a given query image is the ratio of the number of retrieval images relevant to the query image and the total number of images retrieved. It is given by the following formula:

Precision=Number of retrieved images relevant to the query imageTotal number of  images retrieved  .

4.3 Recall

The recall for a given query image is the ratio of the number of retrieval images relevant to the query image and the total number of relevant images in the database. It is given by the following formula:

Recall=Number of retrieved images relevant to the query imageTotal number of relevant images in the database.

4.4 F-measure

A measure that combines the precision and recall into a harmonic way is called F-measure. It is given by the following formula:

F-measure=2PrecisionrecallPrecision+recall.

Table 1 describes the evaluation metrics for the proposed methodology.

Table 1:

Evaluation Metrics.

PrecisionRecallF-measure
0.980.907410.9423
0.960.640.7680
0.90.920.9099
0.860.970.9117
0.790.90.8414
0.870.810.8389
0.910.790.8458
0.90.980.9383
0.840.890.8643
0.960.8670.9111

Here, Table 1 and Figures 79 represent the evaluation metrics for the proposed CBIR system. From these table and figures, we observe that our classification-based retrieval system obtains better accuracy and retrieval rate. This is because the existing CBIR system only concentrates on the retrieval process based on the features extracted. The time to retrieve the image from the huge database that contains images in different modalities is very high for the existing systems. Moreover, it will not extract all the similar images. This issue is solved in our proposed methodology by employing the classification process.

Figure 7: F-measure Evaluation.

Figure 7:

F-measure Evaluation.

Figure 8: Evaluation of Precision.

Figure 8:

Evaluation of Precision.

Figure 9: Evaluation of Recall.

Figure 9:

Evaluation of Recall.

Hence, the proposed CBIR system efficiently retrieves images similar to a given query image based on the extracted features and classification process.

4.5 Comparison

We can begin that our projected work aids to accomplish fine accuracy for the CBIR forecast of input images such as MRI, CT, and ultrasound by using firefly-neural network. Also, we can create this performance of the proposed technique by comparing with other classifiers such as the fuzzy system. Here, we have compared our optimal neural network with fuzzy in which our proposed technique obtained better results.

Figure 10 explains the F-measures for the given input images MRI, CT, and ultrasound. When we compared our proposed F-measure values to the existing technique, our proposed firefly-neural network gave better results, such as 95.35, 95.66, and 96.428; the values of the existing technique are 91.685, 92.514, and 94.3222 (Table 2).

Figure 10: Comparison of Evaluation Measures for F-Measures.

Figure 10:

Comparison of Evaluation Measures for F-Measures.

Table 2:

Comparison of Proposed and Existing F-Measures.

F-measureMRI imagesCT imagesUltrasound images
Existing91.68592.51494.3222
Proposed95.3595.6696.428

The enhanced fine accuracy results of CBIR image retrieval forecast were obtainable by our projected work (Figure 11, Table 3). In contrast, the existing technique provided very low accuracy values for the assessment metrics. The accuracy obtained for our proposed MRI input image for the firefly-neural network was 95.25%, which is high when we compare this value to that of the existing technique (94.5%). In the next input, the CT image obtained 96.99% accuracy, whereas the existing technique obtained 90.35%. In the final input, the ultrasound image obtained 97.54% accuracy, whereas the existing technique obtained 95.25%. From these results, it is acknowledged that our technique performed better.

Figure 11: Comparison of Evaluation Measures for Accuracy.

Figure 11:

Comparison of Evaluation Measures for Accuracy.

Table 3:

Comparison of Proposed and Existing Accuracy Measures.

AccuracyMRI imagesCT imagesUltrasound images
Existing94.595.295.25
Proposed95.2596.9997.54

5 Conclusion

In this paper, we presented an effective multiclass CBIR system. This system is mainly based on three steps, which are segmentation, clustering, and classification. The segmentation process was carried out using the HCBS technique, and it is based on visual features. The segmented images were clustered using the kernel-based fuzzy C-means (KFCM) algorithm. KFCM adopts a new kernel-induced metric in the data space to replace the original Euclidean norm metric in FCM, and the clustered prototypes still lie in the data space so that the clustering results can directly be reformulated and interpreted in the original space. The clustered images were taken as input for firefly-based ANN. ANN classified the clustered images in the appropriate form. This classification led to filtering out the irrelevant images from the database and reduced the search space for further retrieval processed. In the second stage, the relevant images were extracted from the reduced database based on the similarity measure. The proposed CBIR technique was assessed by querying different images, and the retrieval efficiency was estimated by determining the precision-recall values for the retrieval results. The efficiency of the proposed CBIR system was compared with existing systems, which have only the retrieval process. It was hypothesized that an automatic categorization of images would likely perform better in the retrieval process. From the comparative analysis, it can be seen that our system effectively retrieved the images with remarkable precision-recall measurements.

In the future, we will further extend the research by employing sketch-based retrieval in which images of a similar category are recovered. By supervising imperative possessions of sketches, they can detect fine-grained dissimilarities of objects, namely pose and iconic pattern.

Bibliography

[1] S. K. Antani, D. Demner-Fushman, J. Li, B. V. Srinivasan and G. R. Thoma, Exploring use of images in clinical articles for decision support in evidence-based medicine, in: Proc. SPIE-IS&T Electronic Imaging, San Jose, CA, 2008.10.1117/12.766778Search in Google Scholar

[2] S. Antani, R. Long and G. R. Thoma, Content-Based Image Retrieval for Large Biomedical Image Archives, pp. 829–833, Medinfo, San Francisco, CA, 2004.Search in Google Scholar

[3] B. Cheng, S. Antani, R. J. Stanley, D. Demner-Fushman and G. R. Thoma, Automatic segmentation of subfigure image panels for multimodal biomedical document retrieval, in: Proceedings of the SPIE, Document Recognition and Retrieval XVII, San Jose, CA, 2011.10.1117/12.873685Search in Google Scholar

[4] D. Demner-Fushman, S. K. Antani and G. R. Thoma, Automatically finding images for clinical decision support, in: Proceedings of Workshop on Data Mining in Medicine, 7th IEEE International Conference on Data Mining, 2007.10.1109/ICDMW.2007.12Search in Google Scholar

[5] D. Demner-Fushman, S. K. Antani, M. Simpson and G. R. Thoma, Combining medical domain ontological knowledge and low-level image features for multimedia indexing, in: Proc. 2nd International “Language Resources for Content-Based Image Retrieval” Workshop (OntoImage 2008), part of 6th Language Resources and Evaluation Conference (LREC 2008), 2008.Search in Google Scholar

[6] T. M. Deserno, S. Antani and R. Long, Ontology of gaps in content-based image retrieval, J. Digit. Imaging22 (2009), 202–215.10.1007/s10278-007-9092-xSearch in Google Scholar PubMed PubMed Central

[7] A. P. Killedar, V. P. Patil and M. S. Borse, Content based image retrieval approach to tumor detection in human brain using magnetic resonance image, in: First International Conference on Recent Trends in Engineering & Technology, March 2012.Search in Google Scholar

[8] R. McDonald, R. S. Winters, C. K. Ankuda, J. A. Murphy, A. E. Rogers, F. Pereira, M. S. Greenblatt and P. S. White, An automated procedure to identify biomedical articles that contain cancer-associated gene variants, Hum. Mutat.27 (2006), 957–964.10.1002/humu.20363Search in Google Scholar PubMed

[9] A. H. Pilevar, CBMIR: content-based image retrieval algorithm for medical image databases, J. Med. Signals Sensors1 (2011), 12–18.10.4103/2228-7477.83460Search in Google Scholar

[10] Torii, M. and H. Liu, Classifier ensemble for biomedical document retrieval, in: CEUR Workshop Proceedings. 2nd International Symposium on Languages in Biology and Medicine, LBM 2007 - Singapore, vol. 319, Singapore, December 2007.Search in Google Scholar

[11] D. You, E. Apostolova, S. Antani, D. Demner-Fushman and G. R. Thoma, Figure content analysis for improved biomedical article retrieval, Coll. Comput. Digit. Media7247 (2009), 724701–10.10.1117/12.805976Search in Google Scholar

[12] D. You, S. Antani, D. Demner-Fushman, Md. M. Rahman, V. Govindaraju and G. R. Thoma, Biomedical article retrieval using multimodal features and image annotations in region-based CBIR, in: Proceedings of the SPIE, Document Recognition and Retrieval XVII, San Jose, CA, 2010.10.1117/12.838973Search in Google Scholar

[13] D. You, S. Antani, D. Demner-Fushman, Md. M. Rahman, V. Govindaraju and G. R. Thoma, Automatic identification of ROI in figure images toward improving hybrid (text and image) biomedical document retrieval, in: Proceedings of SPIE Electronic Imaging Science and Technology, Document Retrieval and Recognition XVIII. San Francisco, CA, 2011.10.1117/12.873434Search in Google Scholar

Received: 2016-8-25
Published Online: 2016-12-9
Published in Print: 2018-3-28

©2018 Walter de Gruyter GmbH, Berlin/Boston

This article is distributed under the terms of the Creative Commons Attribution Non-Commercial License, which permits unrestricted non-commercial use, distribution, and reproduction in any medium, provided the original work is properly cited.

Scroll Up Arrow