Keywords

1 Introduction

One can offer a tremendous contribution to facilitate doctors and radiologists in (their) work of clinical diagnosis by designing an approach (structure) that could automate their tasks instead of alternately letting them perform manual observations. It gets quite tedious and also incurs surplus time for medical practitioners to manually perform image interpretation. Hence, new vogues for processing of images automatically through computers and medical systems and classifying those images are welcomed now-a-days. Benefitting this needy trend of therapeutic systems, this chore concentrates on developing a methodology (or system) to automate categorization (classification) of X-ray images into six groups. These six groups of X-ray images considered in this study are chest, foot, spine, neck, skull, palm. Images considered in this study are taken from IRMA image CLEF database. This thesis comprises of-four fragments. Fragment 2 provides some information on existing trends, information on work proposed is summarized in Fragment 3, Fragment 4 shows performance assessment with respect to current developed system. Finally, experimental outcomes are handed over to Fragment 5. Fragment 6 gives conclusion.

2 Existing Trends

Enormous work has been done in past and till date by experts across globe in developing medical systems to aid and automate medical diagnosis wherein [1] introduces an effort to interpret digital images, improve diagnosis quality using techniques like median filter, histogram Equalization, for quality enhancement, PCA, K-nearest-neighbour techniques, were applied for selecting features and classification [2]. Introduces a case-study wherein analysis of malicious software is done based on machine learning approaches. Reference [3] introduces a work wherein an effort has been made to develop a CBIR system to retrieve mammographic images of breast tissue and classify them as dense, fatty, glandular using statistical features and SVM. Reference [4] introduces research on a similar area (with splitting/merging scheme) wherein contrast equalizing histogram, FD, ZM and classifiers like KNN, MLP are used. Reference [5] presents work on TRUS images which uses techniques like M3 filter, DBSCAN clustering, SVM to demonstrate a close curative system. Reference [6] presents a work of similar kind wherein classification of images (breast mammograms) based on breast mass is proposed in which segmentation is performed using fuzzy C-means, GLCM features are taken out and adaboost, back propagation, neural networks, sparse representation classifiers are used. Reference [7] presents work wherein pectoral breast muscle cropping is done where Haralicks’ and Zernike descriptors facilitate extraction of features. Classifiers like BPN with SVM operate which resulted accuracy of 95.83%. Reference [8] presents a machine learning based approach to detect/identify images with glaucoma. A series of preprocessing and morphological operations are applied followed by segmenting optic-cup part and extracting rim-disc/cup-disk ratios as features and finally classifying using SVM. Reference [9] In this abstract, collection of (Lung) images are Pre-processed and manipulated with respect to pixel values to either remove distortion, filter the image, enhance contrast for making images fit for better visualization and interpretation [10]. Automated categorizing (of microorganisms) is suggested using SVM classifier. Before performing classification, system feeds image (to module) for Pre-processing then a phase to extract features. Reference [11] introduces thesis presenting a methodology for, classification of lung images in which lung data (images) are collected via lung database with a motto to group them as cancerous/non-cancerous. Prior to the final conclusion, images undergo a set of manipulations constituting Pre-processing (median filter), segmenting (fuzzy c means), GLCM and lastly SVM. Reference [12] shows a machine learning approach to recognize symbols (mathematical symbols) which are handwritten. Reference [13] shows an implementation of a system to detect expressions from a real time dataset of faces by making use of SVM for classification of facial emotions into six categories [14]. Proposes a detection system to ascertain type of breast tissue using GLCM features and FLDA classifier that yielded an accuracy of 72.93% using features based on texture and 82.48% using cascade features [15]. This paper gives an overview on different aspects/studies in educational data mining [16]. Introduces a survey/study on Detection of outliers based on machine learning approaches by considering IOT data for analysis. Reference [17] presents a case study on spam detection based on machine-learning approaches.

Reference [18] Presents classification implementation wherein an approach to classify X-rays by applying M3-filter (better image quality), CCL for segmenting, attributes related to texture, shape are withdrawn by operating with GLCM then classification via SVM [19]. Presents a work on image fusion using guided filters [20]. Introduces a review/study on multi-focus-mage fusion using diversified mechanisms.

3 Proposed Trend

Medical (diagnostic) images accumulated from IRMA (medical image dataset) are fed to Pre-processing module wherein CLAHE (Contrast Limiting AHE (adaptive histogram equalization)) algorithm enhances image calibre (Besides this, other filtering methods like median filter, histogram Equalization; are used to examine, compare the overall system performance) accompanied by segmentation utilizing Connected component Labelling. Texture features for interested region (ROI) are pulled out using GLCM statistics. SVM model is built to which images are fed for classification. System architecture is pictorially introduced in Fig. 1.

Fig. 1
A flowchart of system architecture. I R M A; preprocessing; segmentation; feature extraction; classification. Query; preprocessing; segmentation; feature extraction; classification of feature vector into the chest, foot palm, neck, spine, and skull. Each step is followed by an arrow.

Architecture of proposed trend

Full size image

3.1 A Module to Improve Quality of Images-Pre-processing

For quality intensification (also termed as image enhancement), intensities (or intensity values, pixel values) of digital X-rays are modified which gives finer visualizing, interpretation and displaying experience. It lowers noise within image. One among the pitfalls of median filter is to pull out any outliers (along) with fine or minute details since its’ hard for it to differentiate (distinguish) between the two. Median rate (value) will very minutely be affected by anything diminutive in size. Histogram Equalization emphasizes (stresses) on finding image’s global contrast that gives too dark, too light images. Since it is not ideal for image enhancement, a variation called CLAHE is implemented here. CLAHE concentrates on finding the local contrast of image areas thereby uniformly distributing image intensities. It also clips off or limits contrast to escape noise amplification if any. Sample images demonstrating Pre-processing strategies (techniques) are shown below in Figs. 2, 3, 4, and 5.

Fig. 2
An X-ray and a screenshot of an application side by side. The X-ray on the left displays an original image of a skull. The screenshot on the right displays an equivalent histogram that plots pixel probability versus pixel intensity. Values are estimated. The pixel intensity is maximum at (10, 4500), (125, 1500), and (175, 1700).

Original (skull X-ray) image

Full size image
Fig. 3
An X-ray and a screenshot of an application side by side. The X-ray on the left displays a pre-processed image of a skull. The screenshot on the right displays an equivalent histogram.

Image (X-ray) after Pre-processing (median filter)

Full size image
Fig. 4
An X-ray and a screenshot of an application side by side. The X-ray on the left displays a pre-processed image of a skull. The screenshot on the right displays an equivalent histogram that plots pixel probability versus pixel intensity. Values are estimated. The pixel intensity is maximum at (20, 5000), (49, 5000), (65, 5000), and (75, 5000).

Image (X-ray) after pre-processing (histogram equalization)

Full size image
Fig. 5
An X-ray and a screenshot of an application side by side. The X-ray on the left displays a pre-processed image of a skull. The screenshot on the right displays an equivalent histogram.

Image (X-ray) after pre-processing (CLAHE)

Full size image

3.2 A Module to Extract the Region of Interest-Segmentation

ROI (segmented part) whose features need to be extracted is found using CCL (Connected Component Labelling). This segmentation mode scans an X-ray, group pixels of it considering pixel connectivity. Pixels inside same group share indistinguishable characteristics. Each group of pixels are assigned a label, largest labelled component is the desired region-of-interest (ROI). Since CCL provides a very efficient, user-friendly way to segment the images thereby providing further ease to calculate and display portions of images based on 4-connectivity or 8-connectivity, this has been used in this work.

3.3 A Module to Calculate Statistics-Feature Extraction

When load (input) to algorithm is bulky to be processed (when image sizes are large enough), this input (images) can be represented as compact (feature vector). Here, a statistical methodology called GLCM (Gray Level co-occurrence matrix) is utilized for withdrawing of textural feature-based particulars of the images. GLCM matrices takes (original image’s) pixel intensity co-occurrences into consideration and forms a matrix. Resultant matrix is operated to discover texture statistics (features). Proposed work presented here dealt with extraction of 22 features altogether, couple of them are: correlation, dissimilarity, contrast, variance, entropy, energy homogeneity, etc.

3.4 A Module that does the Purpose-Classification

Feature sets of X-rays (used here) computed in previous segment are forwarded to establish a classifier (SVM here). Since all six categories/classes of images are finely pre-processed and features are withdrawn using ROI, SVM here analyzes these feature sets (of images) and classifies them into six categories where images belonging to same class share similar feature values. SVM principal is depicted pictorially. See Fig. 6.

Fig. 6
A diagram displays a line overlapping on 3 parallel lines. The line at the center and the overlapping line are labeled optimal hyperplane l and separator hyperplane, respectively. The distance between the first and third line is labeled maximum margin. Square box and the circles on the right are labeled classe 1 and 2 support vectors, respectively.

SVM principal

Full size image

4 Performance-Evaluation

Thankfully, Machine Learning (ML) approaches/algorithms are furnished with innumerable alternatives to calculate correctness (accuracy) of any model being developed. This study aims at calculating accuracy (how well did the classifier performed) of classifier by utilizing a confusion matrix (which is a-combination of True positives (TP), True negatives (TN), False positives (FP),False negatives (FN) of predicted versus actual task results. Here the classifiers’ quality is ascertained using specificity, accuracy, sensitivity and specificity measures.

$${\text{Sensitivity}} = {\text{ TP}}/\left( {{\text{TP}}\, + \,{\text{FN}}} \right)\,{\text{Specificity }} = {\text{TN}}/\left( {{\text{TN}}\, + \,{\text{FP}}} \right)\,{\text{Precision}}\, = \,{\text{ TP}}/\left( {{\text{TP}}\, + \,{\text{FP}}} \right)$$
$${\text{Accuracy}} = \left( {{\text{TP}}\, + \,{\text{TN}}} \right)/\left( {{\text{TP}}\, + \,{\text{TN}}\, + \,{\text{FP}}\, + \,{\text{FN}}} \right)$$

5 Investigational Outcomes

Eventually, CLAHE Pre-processing worked best with GLCM, SVM offering 96% accuracy unlike median filtering, histogram filtering wherein classification result or accuracy was found as 91% ,89% respectively. Overall performance results of classifying X-ray images preprocessed using CLAHE are tabulated in Table 1 wherein; for the class Skull, precision, sensitivity, specificity values are 100%, 80%, 100% respectively with an accuracy of about 96.55% in contrast to the class Foot wherein the accuracy is much lesser i.e.; about 89.65%.Overall performance results of classifying X-ray images preprocessed using Histogram Equalization are tabulated in Table 2 wherein; for the class Neck, precision, sensitivity, specificity values are 100%, 70%, 100% respectively with an accuracy of about 94.82% in contrast to the class Spine wherein the accuracy is much lesser i.e.; about 84.48%.Overall performance results of classifying X-ray images preprocessed using Median Filter are tabulated in Table 3 wherein; for the class Neck, precision, sensitivity, specificity values are 100%, 80%, 100% respectively with an accuracy of about 96.55% in contrast to the class Spine wherein the accuracy is much lesser i.e.; about 84.48%. Figures 7, 8 and 9 presents output X-rays (images) post pre-processing, segmentation when those images were preprocessed using CLAHE, Histogram Equalization and Median Filter respectively. The performance measures discloses (shows) that CLAHE preprocessed (X-ray) along with GLCM (texture features) and SVM (classifier) gives a better/finer accuracy over other filters and serves the purpose. Hence, SVM-classification is quite acceptable for classifying X-rays (images). A bar graph (as) in Fig. 10 is opted to present a pictorial/vivid look on classification results.

Table 1 Performance results (measures) for GLCM, SVM- with CLAHE
Full size table
Table 2 Performance results (measures) for GLCM, SVM -with histogram equalization.
Full size table
Table 3 Performance results (measures) for GLCM, SVM with- median filter
Full size table
Fig. 7
A chart of X-rays in six rows and four columns. The column headers are input, pre-processed images, C C L, and segmented images of the chest, foot, hand, face, neck, and spine from rows 1 to 6, respectively.

Represents X-rays post CLAHE

Full size image
Fig. 8
A chart of x-rays in four columns and six rows. The column headers are input, pre-processed images, C C L, and segmented images. The row entries from columns 1 to 4 are as follows. Row 1: chest; row 2: foot; row 3: hand; row 4: face; row 5: neck; and row 6: spine.

Represents X-rays post histogram equalization

Full size image
Fig. 9
A chart of X-rays in four columns and six rows. The column headers are input, pre-processed images, C C L, and segmented images. The row entries from columns 1 to 4 are as follows. Row 1: chest; row 2: foot; row 3: hand; row 4: face; row 5: neck; and row 6: spine.

Represents X-rays post median-filter

Full size image
Fig. 10
A grouped bar graph. Values are estimated. Median filter, G L C M, and S V M (precision, 84), (sensitivity, 63), (specificity, 97), (accuracy, 91). Histogram E Q, G L C M, S V M (precision, 73), (sensitivity, 62), (specificity, 96), (accuracy, 89). Clahe, G L C M, and S V M (precision, 92), (sensitivity, 82), (specificity, 98), and (accuracy, 95).

Performance results (measures) for overall percentage of classified X-ray image

Full size image

6 Conclusion

In Proposal described here, medical (X-rays) images of six categories namely chest, spine, palm, skull, foot, neck are acquired from IRMA medical database. These images (X-rays) are subjected to Pre-processing module wherein three different image manipulation techniques/strategies namely Median-filter, Histogram Equalization and CLAHE are implemented (or worked with) to compare their performance/results and find which Pre-processing technique works best with GLCM, SVM. Images when Pre-processed with CLAHE outshined among all three with accuracy of 96% altogether, sensitivity of 82.91%, specificity of 97.96% and precision of 92.91%. This gave a conclusion that images when Pre-processed with CLAHE gave better accuracy post classification. Future scope of this proposed work is the extension (of current work) for content based image retrieval systems (CBIR). Unlike the six groups/classes of X-rays images specified above, other parts like knee, elbow, limbs, ankle etc. can also be taken into consideration for future upgrading of medical systems. A Heterogeneous (image) types like C.T, M.R.I, U.S.G etc. can also be worked with in future.