Auto-shape Lossless Compression of Pharynx and Esophagus Fluoroscopic Images

J Med Syst (2015) 39:5 DOI 10.1007/s10916-015-0200-z

SYSTEMS-LEVEL QUALITY IMPROVEMENT

Auto-shape Lossless Compression of Pharynx and Esophagus Fluoroscopic Images Arif Sameh Arif & Sarina Mansor & Rajasvaran Logeswaran & Hezerul Abdul Karim

Received: 29 September 2014 / Accepted: 13 January 2015 # Springer Science+Business Media New York 2015

Abstract The massive number of medical images produced by fluoroscopic and other conventional diagnostic imaging devices demand a considerable amount of space for data storage. This paper proposes an effective method for lossless compression of fluoroscopic images. The main contribution in this paper is the extraction of the regions of interest (ROI) in fluoroscopic images using appropriate shapes. The extracted ROI is then effectively compressed using customized correlation and the combination of Run Length and Huffman coding, to increase compression ratio. The experimental results achieved show that the proposed method is able to improve the compression ratio by 400 % as compared to that of traditional methods.

Keywords Fluoroscopic medical images . ROI . Lossless image compression . Run-length Coding . Huffman coding . Correlation

This article is part of the Topical Collection on Systems-Level Quality Improvement A. S. Arif : S. Mansor (*) : R. Logeswaran : H. A. Karim Faculty of Engineering, Multimedia University, 63100 Cyberjaya, Selangor, Malaysia e-mail: [email protected] A. S. Arif Foundation of Technical Education, Institute of Administration Rassafa, Baghdad, Iraq R. Logeswaran Asia Pacific University of Technology & Innovation, 57000 Kuala Lumpur, Malaysia

Introduction Many techniques have been developed to enhance medical diagnoses. The most widespread is the use of medical imaging. The amount of data collected and stored increases costs. In spite of the developments in modern medical imaging techniques, or rather because of the ability in acquiring ever higher resolution images, medical institutions are generating larger medical image databases. Hospitals and medical centers are witnessing an increase in the numbers and types of such images, which leads to a huge number of image series from different examinations for different cases. The most common technique to increase the storing and exchanging efficiency is by decreasing the physical size of images. Greater benefits are gained from compressing the size of series of images rather than a single image [1]. Due to the need for accurate diagnosis, lossless image compression is the crucial approach in reducing the undesired size without loss in the authenticity of the images, allowing the images to be completely restored upon decompression [2, 3]. However, the best approach to gain higher compression performance is through lossy compression, which risks losing potential important details in the decompressed image. Some studies such as [4] have highlighted the advantages of lossy compression in medical diagnosis in order to legitimize it. They mentioned that based on the noisy nature of medical imaging equipment, medical inspectors preferred controlled lossy compression for medical images. This conclusion was derived after a double-blind test with radiologists, to gage the accuracy of the tested images [3].

5

J Med Syst (2015) 39:5

Page 2 of 7

Background Fluoroscopy is a type or branch of X-ray imaging that produces multiple X-ray images of the organ or structure being examined. It is a common tool in many types of inspections and techniques, such as cardiac catheterization, percutaneous Nephrostomy (PCN), and Barium Sallow. This type of imaging provides simultaneous assisting information to doctors during surgery [5, 6]. Region of interest (ROI) represents the most important segment in an image. Focusing on ROI reduces the cost of time and processing for irreverent segments of the image. Extracting an ROI from medical images is a crucial task as it is the important part of the images that should be focused on first. In transmission-based ROI coding methods, the coefficients associated with the background are transmitted after those associated with the ROI. Therefore, it is necessary to identify the coefficients required for the reconstruction of the ROI. The ROI mask determines which coefficients have to be transmitted to reconstruct the ROI at the receiver [7, 8]. Another challenging work was to design a technique that can deal with the common kinds of medical images such as Xray, MRI, and ultrasound images [9]. They focused on the diagnostically important regions using segmentation technique, self-organization map (SOM). The two-stage SOM was utilised to identify dominant colours of the image precisely, then the image is segmented into several small segments. Recursive merging steps were appropriately conducted to merge smaller regions into larger ones in order to identify ROI’s by the radiologist. An efficient method was presented to improve the trade-off between edge detection and noise reduction [10]. Tumour region is detected in brain MRI image automatically; then they use compass operator depending on 4th order derivative as noise reduction technique. A new morphological technique is also introduced for stripping skull region from the brain images, which consequently leading to the process of detecting tumour accurately. ROI regions were detected in this work using automatic seeded segment growing technique; interesting results were yielded in this work. Image compression techniques utilise coding algorithms in order to limit the length of the coded image. Huffman coding is a technique that assigns variable length codes to yield the best size for the coded message. It assigns the shortest code word to the most frequent element, which specifies the optimal size. Consequently, longer code words are assigned to elements with lower frequencies [11]. The authors proposed a lossless compression technique for CT medical images [12]. Image data were clustered into two streams; first one represents the positions of the remaining data after performing the segmentation. They use JBIG compression technique to compress selected data of this stream.

Second stream contains real values of data, which will be compressed using lossless compression technique. Another well-known coding technique is Run-length Coding; the main idea of this technique is to utilise the successive values and replace a long series of repeated values with two values: the essential value and the number (count) of these values [13]. The combination of Run-length and Huffman coding techniques have been found to produce more efficient compression results by minimizing the data size and pattern delivery time, in addition to saving processing abilities in scan applications [14]. For medical image compression, combinations of Run-length and Huffman techniques were efficiently applied, especially on angiograms, X-ray, and MRI images, in order to maximize compression results [15]. In previous works, we developed a lossless compression technique. It was applied on Fluoroscopic medical images using Huffman coding for compression and correlation for determining the relation between an image and the containing set [16]. The average compression ratio achieved by the proposed method was 7.97. An extended work was applied to utilise the combination of Huffman and Run-length compression techniques. The developed technique achieved an improved compression ratio of 11.41 [17]. We then improved our technique by extracting the ROI from each image and encoding the difference ROI with the combination of finding the differences between the two ROIs, Run-length and Huffman coding (R. Huff-Seg.). In our latest work [18], extracting the ROI contributed to a major improvement in compression ratio, the developed method reached significantly improved compression ratio of 29.73. However, we only used the quadrant shape to fit the ROI, which is not fully optimized. In this paper, we propose the use of multiple shapes in extracting ROIs. An automated method is developed to identify the suitable shape for each ROI, with the aim of further improving the compression performance. To the best of our knowledge, there are no other published works employing this technique in the context of medical image compression. Proposed framework The proposed frame work consists of four main stages: Classification, Segmentation, Subtraction and Coding. By calculating the Correlation Coefficient (CC), the set of images for each patient can be classified into groups based on the view/ orientation of capture. The background and implementation details of the classification stage can be found in our previous work in [16]. The focus of this paper is on the second stage, as it is the segmentation that tries to omit the diagnostically insignificant details from the image [17]. More specifically, this paper discusses an improved extraction mechanism for ROI. The

J Med Syst (2015) 39:5

Page 3 of 7 5

approach will be established in the next section. The third stage, subtraction generates the difference vector between the test image and the reference image [18]. The final stage will be to encode the difference vector in a more compressed form by applying the developed combination of Run-length and Huffman coding [19].

ROI

Methodology Fluoroscopy images can be divided into three regions as shown in Fig. 1: & & &

Background regions, which consist of four black angles and two white half circles. Region of Coding (RC), which is the rectangular area that contains the image of the patient’s oesophagus and pharynx. Region of interest (ROI), which is the smallest area that contains the oesophagus and pharynx.

The RC consists of two regions, as shown in Fig. 2, namely: 1- Region of Interest (ROI); 2- Background Region of Interest (B-ROI).

B-ROI Fig. 2 Important areas in Region Coding (RC) Fluoroscopic images

extracted, as in Fig. 3(e). The steps are well demonstrated in the previous work in [17]. A summary of the steps involved is as follows: & & & & &

This work tries to minimize the redundancy resulting from ROI extracted using the ordinary quadrant shape. The ROI shape is anatomized in order to choose the minimum shape and size that contains the ROI with the least number of nonROI pixels; this is called Auto Shape.

& & &

Histogram analysis to identify the peaks in the image (see Fig. 3b). Threshold specified based on the peaks in the histogram (see Fig. 3c for thresholded image). Calculation of the areas inside the image to specify the pharynx and the esophagus. Generation of an ROI mask. Application of the combination of (Bilat_Aniso) filter to smooth the edges [20]. Application of a median filter and the morphological erosion technique to smoothen the mask. Segmentation of the thresholded image (see Fig. 3d). Exracting the ROI by applying the label image on the original image to extract the ROI. Fig. 3(e) shows the ROI image.

Extract the ROI In this work, where the ROI is the esophagus and pharynx, there is only one part of the image that needs to be compressed losslessly. This area would need to be segmented and

Black area White area RC Fig. 1 Important areas in Fluoroscopic images

Auto shape ROI After extracting the ROI, the area surrounding the ROI is minimized by cropping the RC to a minimum quadratic shape containing the ROI. This simplistic approach yields redundant areas that could be further compressed. Sample images (a) and (d) with their corresponding quadrant cropped images ((b) and (e)) and indication of the remaining redundant areas surrounding the ROI that are to be compressed ((c) and (f)), are shown in Fig. 4. To overcome the redundancy in this cropping method, more suitable shapes that are able to contain the ROI with minimum presence of redundant pixels should be used. The chosen shape must contain the ROIs in both reference and test images. For example, the same parallelogram shape would be

5

J Med Syst (2015) 39:5

Page 4 of 7

Fig. 3 ROI extraction steps

The highest peak represents the threshold value of the edges. (a) Original image

(b) Intensity his ogram & identified threshold value

c) Threshold image (d) Labeled image

used for the ROIs in both images during ROI extraction, as shown in Fig. 5. The choice of a suitable shape for the ROI depends on the labeled image, where necessary information such as the shape type and its coordinates are extracted. Using this information, the chosen shape containing the ROI is cropped from the

(a)

(b)

(c) Redandannt

(e) ROI image

original images (see Fig. 6). The same information is necessary for reconstructing the shape upon the decompression stage. In this work, for the organ of interest, the suitable shapes identified include rectangle, square, triangle, parallelogram and circle, as shown in Fig. 7. The dimensions of the shape, such as the width, height and/or diameter, would depend on the identified coordinates. Compression of the ROI difference between the two images (original and reference) within the chosen shape will be discussed in the next stage. Some vital information is added to the output (transmission stream or file, depending on the application) of the compression for the purpose of reconstruction. This information is in terms of an indicator of the shape type and the required coordinates that represents both the position and dimensions of the shape. In addition to the compression output, reference image will be stored in order to contain the reconstructed shape.

area ROI

Redundant R arreas

(d)

(e)

Fig. 4 Remaining redundancy inside the segmented images

(f)

iggnored

Fig. 5 ROI and redundant area

J Med Syst (2015) 39:5

Page 5 of 7 5

Fig. 6 ROI on the original images Fig. 8 Reconstruction process

Reconstruction can be performed by analyzing the compression output and extracting the necessary indicators. A summary of the steps taken during reconstruction is as follows: 1. Reverse the encoding process of the compression algorithm [18] to produce a data stream. 2. Extract the first value in the stream (which is the shape type indicator), followed by the coordinates. 3. Extract the reference image, which is the first image in the stream. It represents the essential complete image of the patient before the liquid (contrast agent) motion (see the middle image in Fig. 8).

4. Using the shape type and coordinates, the position of auto shape ROI area on the reference image is identified. 5. Extract the remaining image data in the stream. Each represents the ROI information within the chosen auto shape of an image (as shown in the first image of Fig. 8). 6. Add the auto shape ROI information to the reference image to produce the reconstructed original image (see the last image in Fig. 8). 7. Repeat step 6 above with each image in the sequence to reconstruct all the images of that sequence.

Encoding

Rectangle

Circle

Triangle

The subtraction produces a one dimensional vector. The method or algorithm used for coding is determined by the construction of the vector data, where the best choice would produce the smallest compressed output. As there is much similarity between successive images in a sequence, the difference vector has the characteristic of containing long runs of repeated small values. The combination of Run-length and Huffman coding is selected because they are based on duplication and the estimated probability of appearance for each possible value of the source symbol, hence most well suited for this purpose.

Results and discussion Experiments were performed on 386 greyscale still clinical fluoroscopic images obtained from Serdang Hospital, Malaysia to evaluate the validity of the proposed approach. Each image was of dimension 512×512 pixels and file size of 256 KB. The performance evaluation, in terms of Compression Ratio (CR), is calculated as follows: CR ¼ where, Parallelogram Fig. 7 Various Autoshapes to suit the ROI

Square

So Sd

So Sd

File size of the original image File size of the difference image.

ð1Þ

5

J Med Syst (2015) 39:5

Page 6 of 7

As each sequence contains many images and the CR performance achieved for each image varies depending on its similarity with the reference image, the average CR achieved for three sequences (groups) of images is calculated and presented in Table 1. From the results for each group as well as the overall average given in the last row, it is obvious that the proposed method achieved significantly better results over the existing methods. Table 2 shows the correlation values between the reference and the tested images for a random sample of 20 images. The higher correlation corresponds to a higher similarity, resulting in a higher CR. From Table 2, it is also observed that the proposed method (Auto shape R. Huff) achieved significantly better performance than implementing Run-length Huffman ROI (R. Huff-ROI) on the difference images and it also performed better than the other two methods, namely, Run-length Huffman on difference images (R. Huff.-Diff.) and the standard lossless combination (R. Huff) compression of the images. For example, in the case of images 005 and 006, the standard combination method (R. Huff) produced a CR of only 1.38. Run-length Huffman coding for the difference images (R. Huff-Diff) realized a CR of 11.41, while (R. Huff-ROI) for the ROI difference images achieved a CR of 12.21. However, the proposed method (Auto shape-R. Huff) implemented based on the CC indications, achieved a much higher CR of 49.01 for the Auto shape ROI difference image (005–006). The improvement in compression happens when suitable shapes are implemented to extract the ROI with a minimum amount of background. To validate the quality of performance of the proposed framework for image compression, the error between the original image and decompressed image was calculated using MSE (Mean Square Error). The results obtained were a perfect match of zero error, proving that the developed technique is lossless.

Table 1 Comparison of the ACR performance between the proposed method (Auto shape-R. Huff) as compared to previous methods on the standard image for a random sample of ten images Group no.

No. of Average of Compression Ratio (ACR) images R. Huff R. Huff-Diff R. Huff-Seg. Auto shape-R. [16] [14] Huff

011 330 954 Average

13 19 25 47

1.35 1.31 1.34 1.33

8.22 9.85 5.85 7.97

16.15 18.83 18.76 17.91

50.10 104.75 70.73 75.19

Table 2 Comparison of CR performance between the proposed method (Auto shape-R. Huff) on the difference segmented images as compared to the previous methods (R. Huff, R. Huff-Seg. and R. HuffDiff) for a random sample of ten images Ref Test CC* Compression Ratio (CR) image image R. R. Huff R. Huff-Seg. Auto Huff (Diff) [16] [14] shape-R. Huff 33005 33004 33009 33007 33010 33006 33012 01104 01108 01105 01112 01111 01106 01107 01113

33006 33005 33010 33008 33011 33007 33013 01105 01109 01106 01113 01112 01107 01108 01114

0.93 0.95 0.95 0.95 0.95 0.95 0.98 0.86 0.94 0.88 0.98 0.95 0.95 0.97 0.97

1.38 1.34 1.36 1.36 1.35 1.36 1.35 1.33 1.32 1.33 1.35 1.37 1.34 1.35 1.36

11.41 11.20 11.04 10.55 9.81 10.79 10.57 5.91 5.88 5.73 6.84 7.11 6.64 6.87 6.8

12.21 14.82 13.83 12.88 12.83 16.51 20.89 9.29 9.27 7.32 13.36 15.54 14.77 16.25 15.41

49.01 46.78 77.65 48.65 79.38 50.67 87.15 37.53 25.76 57.13 16.33 19.52 24.04 35.13 46.38

Conclusion In this paper, a new framework for lossless image compression of groups of images, based on their correlation, has been proposed. The two main contributions of this work are: (1) an improvement in identifying the outline shape of ROI, and (2) concentration on electing the minimum area that represents the most important parts of the fluoroscopic images. The technique concentrates on consideration of automatically choosing the most appropriate shape from a selection of shapes, to contain the two ROI regions (i.e., in the reference and test images) with minimum redundant areas present in the extracted ROI regions. Subsequently, the difference between the groups of images is encoded using a combination of Runlength and Huffman coding. From the results obtained, this lossless method is able to achieve significant improvements in compression performance, with benefits in storage and transmission.

Acknowledgments The authors would like to thank Dr. Sharifah Mastura Syed Abu Bakar (Head Radiology Department), Khatijah Ali (Radiographer) and Mr Ang Kim Liong (Clinical Research Centre) at Serdang Hospital, Malaysia, for their assistance and collaboration in undertaking this work.

J Med Syst (2015) 39:5

References 1. Bairagi, V., and Sapkal, A., Automated region-based hybrid compression for digital imaging and communications in medicine magnetic resonance imaging images for telemedicine applications. IET Sci. Meas. Technol. 6(4):247–253, 2012. 2. Kumar, N., and Kamargaonkar, C., A survey on various medical image compression techniques. Int. J. Sci. Eng. Technol. Res. 2(2): 501–506, 2013. 3. Ukrit, M., Umamageswari, A., and Suresh, G., A survey on lossless compression for medical images. Int. J. Comput. Appl. 31(8):47–50, 2011. 4. Choong, M., Logeswaran, R., and Bister, M., Improving diagnostic quality of MR images through controlled lossy compression using SPIHT. J. Med. Syst. 30(3):139–143, 2006. 5. Mei-Yen, C., Chien-Tsai, L., Ying-Chou, S., et al, Design and evaluation of a DICOM compliant video fluoroscopy imaging system. In: Proceeding of IEEE 9th International Conference on Health Networking Application and Services: Taipei, pp. 248–251, 2007. 6. Russakoff, D., Rohlfing, T., and Maurer, C., Fuzzy segmentation of X-ray fluoroscopy images. Proc. SPIE 4684:146–154, 2002. 7. Doukas, C., and Maglogiannis, I., Region of interest coding techniques for medical image compression. IEEE Eng. Med. Biol. Mag. 26(5):29–35, 2007. 8. Zhang, Q., Xiao, H., Extracting regions of interests in biomedical images. IEEE International Seminar on Future Biomedical Information Engineering, pp. 3–6, 2008. 9. Teng, W., and Chang, P., Identifying regions of interest of medical images using self-organizing map. J. Med. Syst. 36:2761–2768, 2011. 10. Faisal, A., Parveen, S., and Badsha, S., Assisted diagnostic systems in tumor radiography. J. Med. Syst. 37:9938, 2013.

Page 7 of 7 5 11. Špelič, D., and Žalik, B., Lossless compression of thresholdsegmented medical images. J. Med. Syst. 36:2349–2357, 2012. 12. Song, T., and Shimamoto, T., Reference frame data compression method for H.264/AVC. IEICE Electron. Expr. 4(3):121–126, 2007. 13. Núñez, J., and Jones, S., Run-length coding extensions for high performance hardware data compression. IEE Proc. Comp. Digit. Tech. 150(6):387–395, 2003. 14. Norani, M., and Tehranipour, M., RL-Huffman encoding for test compression and power reduction in scan applications. ACM Trans. 10(1):91–115, 2005. 15. Sunder, R., Eswaran, C., Sriraam, N., Performance evaluation of 3-D transforms for medical image compression. IEEE International Conference, Electro Information Technology, 6 pp. -6, 2005. 16. Arif, A., Mansor, S., Logeswaran, R., and Abdul Karim, H., Lossless compression of fluoroscopy medical images using correlation. J. Asian Sci. Res. 11(2):718–723, 2012. 17. Sameh, A., Mansor, S., Logeswaran, R., Abdul Karim, H., Segmentation and compression of pharynx and esophagus fluoroscopic images. IEEE International Conference on Signal and Image Processing Applications (ICSIPA), 2013. 18. Arif, A., Mansor, S., Logeswaran, R., and Abdul Karim, H., Lossless compression of pharynx and esophagus in fluoroscopy medical images. Int. J. Biosci. Biochem. Bioinforma. 3(5): 483–487, 2013. 19. Arif, A., Mansor, S., Logeswaran, R., Abdul Karim, H., Lossless compression of fluoroscopy medical images using correlation and the combination of Run Length and Huffman coding. IEEE International Conference on Biomedical Engineering & Science (IECBES), pp. 759–762, 2012. 20. Arif, A., Mansor, S., Logeswaran, R., Abdul Karim, H., Combined bilateral and anisotropic-diffusion filters for medical image denoising. IEEE Student Conference for research and development (SCOReD), pp. 420–424, 2011.