Scalable Medical Data Compression and Transmission Using Wavelet Transform for Telemedicine Applications

54 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 7, NO. 1, MARCH 2003 Scalable Medical Data Compression and Transmission Using Wavelet Transform for Telemedicine Applications Wen-Jyi Hwang, Associate Member, IEEE, Ching-Fung Chine, and Kuo-Jung Li Abstract In this paper, a novel medical data compression algorithm, termed layered set partitioning in hierarchical trees (LSPIHT) algorithm, is presented for telemedicine applications. In the LSPIHT, the encoded bit streams are divided into a number of layers for transmission and reconstruction. Starting from the base layer, by accumulating bit streams up to different enhancement layers, we can reconstruct medical data with various signal-tonoise ratios (SNRs) and/or resolutions. Receivers with distinct specifications can then share the same source encoder to reduce the complexity of telecommunication networks for telemedicine applications. Numerical results show that, besides having low network complexity, the LSPIHT attains better rate-distortion performance as compared with other algorithms for encoding medical data. Index Terms Image coding, telemedical communication network, wavelet transform. I. INTRODUCTION THE term telemedicine refers to the use of communications technology to deliver medical care without regard to the distance that separates the participants. With the rapid development of communications technology, the applications and services of telemedicine are growing [5]. A comprehensive system integrating various applications within a common infrastructure is usually necessary. This infrastructure includes the physical facilities and equipment used to capture, transmit, store, process, and display medical data and images. To effectively reduce the cost of realizing the infrastructure, data compression is usually desired so that the bandwidth and storage requirements for delivering and saving medical data are lowered. The equipment of a telemedicine system may vary widely because of the cost-effectiveness considerations of different applications. Different displaying/processing devices, based on their capacities, may have distinct demands on the information delivered from a transmitter or source encoder. Those demands may be specified in terms of the signal-to-noise ratio (SNR) and/or resolution of the reconstructed data. Because of the variety of requests from this equipment, it is difficult for the transmitter to provide an encoded bit stream satisfying all the requirements. One solution to this problem is to use the simulcast technique where medical data are encoded independently for each of the devices. The transmitter will then deliver the The authors are with the Department of Electrical Engineering, Chung Yuan Christian University, Chungli, 32023, Taiwan (email: whwang@ dec.ee.cycu.edu.tw). Digital Object Identifier 10.1109/TITB.2003.808499 encoded bit stream based on the specific user request. This approach requires more resources to be used in the encoder in terms of disk space and management overhead. Therefore, a scalable transmission system is usually desired. In the scalable transmission scheme, an encoded bit stream is delivered in two or more layers. At the receiving ends, medical data in different SNRs and/or resolutions are reconstructed by decoding bit streams accumulated up to different enhancement layers from the base layer. Users with various SNR and resolution requirements therefore can share the same source encoder and transmission system. Although many algorithms are effective for medical data compression [4], they can not perform scalable coding. The wavelet-based embedded coding algorithms [3], [7], [8] can be employed for realizing scalable systems. Some of these algorithms, such as the set partitioning in hierarchical trees (SPIHT) [6], [7] algorithm and JPEG2000 [8] have been found to outperform many existing methods for image and/or electrocardiogram (ECG) compression. However, the SPIHT algorithm is only SNR scalable since it always encodes and displays reconstructed data at a fixed resolution level (usually the full resolution). The JPEG2000 can be used for SNR or resolution scalable transmission. The algorithm allows the compression ratio (CR) at each layer to be pre-specified. Therefore, the SNR scalability can be attained if the bit stream is delivered in layer-resolution-component-position order. Nevertheless, each layer of the JPEG2000-encoded bit stream covers all the subbands of the medical data. The resolution scalability, therefore, can not be achieved in this delivery order because the medical data is always displayed at full resolution for all layers. The resolution scalable transmission can be realized by setting the transmission order to resolution-layer-component-position or resolution-position-component-layer, where the low-pass subband at each resolution level is displayed progressively. The JPEG2000 uses the post-compression rate-distortion optimization (PCRD-opt) [8] algorithm for determining the rate allocated to each subband. Therefore, the rate used for the encoding of the low-pass subband at each resolution level can not be pre-specified/controlled. That is, the resolution and CR can not be controlled independently in the JPEG2000 scheme. This paper employs a layered SPIHT (LSPIHT) technique for the design of scalable transmission systems. In the LSPIHT, the CR and resolution associated with each layer can be pre-specified before encoding. The transmitted images are reconstructed with CR and resolution identical to those of the highest layer 1089-7771/03$17.00 2003 IEEE

HWANG et al.: SCALABLE MEDICAL DATA COMPRESSION AND TRANSMISSION 55 Fig. 2. Structure of the QMF for the decomposition of x into four subbands x, x, x, x, where g and h denote the analysis filters. Fig. 1. Wavelet transform coefficients of an image x. accumulated by the decoder. Both the SNR and resolution scalabilities therefore can be achieved. To satisfy both CR and resolution constraints at each layer, starting from the base layer, the LSPIHT encodes one layer at a time until the design of the top layer is completed. The encoding of each layer is based on SPIHT which only covers the subbands with resolution lower than the resolution constraint of that layer. To enhance the performance of SPIHT at each layer, the encoding results of the previous layers are used. In this paper, the LSPIHT is applied to ECGs and magnetic resonance images (MRIs) for scalable transmission, where different layers are associated with distinct CRs and resolutions. Simulation results show that, as compared with other data compression techniques, the LSPIHT algorithm attains better rate-distortion performance for the encoding of each layer while demanding fewer resources and lower costs. II. WAVELET TRANSFORM, SIMULCAST AND SCALABLE TRANSMISSION SYSTEMS Since the LSPIHT algorithm is based on a wavelet transform, this section first briefly reviews some basic facts of the discrete wavelet transform (DWT). After that, the basic structure of the simulcast and scalable transmission systems are described. Although this paper considers the compression of both onedimensional (1-D) and two-dimensional (2-D) medical data, we focus the 2-D LSPIHT algorithm since 1-D LSPIHT is a simple degenerate version of 2-D LSPIHT. Therefore, we first recall the definition of 2-D DWT for 2-D LSPIHT. Let be an image to be transmitted over the scalable system. The dimension of is assumed to be 2 2. Let be the -stage wavelet transform matrix of, where. Then, as shown in Fig. 1, is also a 2 2 matrix containing subbands and,,,, each with dimension 2 2. Note that, in the wavelet transform matrix, the subbands (low-pass subbands at resolution level ), and,, (, and orientation selective high-pass subbands at resolution level ),, are obtained recursively from Fig. 3. Example of wavelet tree. with, where the resolution level is also referred to as the full resolution. The decomposition of into four subbands,,, can be carried out using a simple quadrature mirror filter (QMF) scheme as shown in Fig. 2 [9]. The wavelet coefficients can be organized as a set of trees, called wavelet trees, for image coding. In the wavelet domain, with the exception of the subbands at lowest resolution level, every coefficient at a given resolution level can be related to a set of coefficients of the same orientation at the next higher resolution level. The coefficient at the low resolution level is called the parent, and all the coefficients at the same spatial location and of the same orientation at the next higher resolution level are called children. For example, Fig. 3 shows a wavelet tree originating from a coefficient in the subband. For the lowest resolution subband,, in this example, the parent-child relationship is defined such that each parent node has three children, one in each subband at the same resolution level and spatial location but different orientation. The basic structure of the simulcast system is shown in Fig. 4(a). In the system, each layer is associated with a CR and a resolution, which can be prespecified independently. Suppose there are layers in the scalable system. Let and be the rate and resolution associated with layer, respectively. The rate

56 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 7, NO. 1, MARCH 2003 (a) (b) Fig. 4. Basic structure of the simulcast and LIT systems: (a) Simulcast system. (b) LIT system. is the number of bits per pixel for storage and transmission at layer, and therefore can be used to compute the CR at layer. The medical data are then encoded independently for each layer subject to the CR and resolution constraints at that layer. Consider an image for data transmission. The objective of encoding process at the layer is to independently code the

HWANG et al.: SCALABLE MEDICAL DATA COMPRESSION AND TRANSMISSION 57 low-pass subband with rate. The decoders therefore reconstruct the medical data in different CR and resolutions by collecting bit streams from different layers. The system is simple to implement. However, the low-pass subbands at different layers are highly correlated. Hence, encoding these low-pass subbands independently may result in large overhead for data transmission. Typical implementation of the scalable transmission systems is based on the structure of the layered image transmission (LIT) system shown in Fig. 4(b), which can be used to eliminate the drawback stated above. In the LIT, the CR and resolution associated with each layer are also desired to be prespecified. In addition, the layers are arranged in such a way that layers having lower resolution are placed in lower positions. That is, for. The low-pass subband is also encoded at layer. However, the encoding process at each layer is not performed independently. Note that is a low-pass subband of when. Hence, the encoder at layer can utilize the encoding results at the previous layers to reduce the overhead for scalable transmission. To reconstruct medical data at each layer, the decoder has to accumulate bit streams up to that layer starting from the base layer. Assume the transmission is lossless. The resolution of the reconstructed medical data after decoding is the resolution of the layer in the highest position among the layers decoded by the receiver. Let the incremental rate at layer of the LIT, denoted by, be the number of bits per pixel at layer. The accumulated rate up to layer, denoted by, can be obtained by where. Similar to in the simulcast system, the incremental rate is used to compute the CR at layer. On the other hand, the accumulated rate indicates the number of bits required for decoding at layer. In the SPIHT and the JPEG2000 algorithms, the rate for encoding an image can be prespecified. Therefore, these algorithms can be applied for the implementation of simulcast systems where the at each layer, are treated as separate images, and are encoded independently with rate.in the LIT systems, however, the,, are viewed as the subbands of. Since the rate allocated to these subbands can not be pre-specified in the SPIHT and JPEG2000 algorithms, it may be difficult to control the incremental rate at each layer using these algorithms for the implementation of the LIT. The LSPIHT algorithm can be used to solve this problem. Detailed description of the algorithm is given in the next section. III. LSPIHT ALGORITHM The LSPIHT algorithm can be viewed as a sequence of operations using the SPIHT algorithm with one operation for each layer. Starting from the base layer, the LSPIHT constructs one layer at a time until the design of the final layer is completed. The objective of the LSPIHT at each layer is to encode using the SPIHT algorithm with the rate budget. Therefore, the encoding region at layer only covers the subbands,,,,. Assume when. Therefore, the encoding region at layer also comprises (1) Fig. 5. Example for the encoding of LIT using SPIHT. the encoding region at the layer,. Consequently, the encoding process at each layer can use the encoding results at its previous layers to reduce the overhead for information delivery over the LIT. Fig. 5 shows a simple example for the design of the LIT using the LSPIHT technique. In Fig. 5, the dimension of an image is assumed to be 2 2 (i.e., ). The number of layers, and resolution levels associated with layer 1 and layer 2 are and, respectively. Because, the subbands,,,,, 1, 2, (which are in the shaded region of Fig. 5) are the subbands constituting layer 1. In addition, since, all the subbands in Fig. 5 (including the subbands in the shaded region) are used for the encoding of layer 2. Therefore, independent encoding of layers 1 and 2 may result in large overhead. By contrast, using the encoding results at layer 1 for the encoding operation at layer 2 can effectively reduce this overhead. In the following we describe the 2-D LSPIHT technique in more detail. Recall that a wavelet coefficient is said to be significant for bit depth if 2, otherwise it is said to be insignificant. Moreover, a wavelet tree is said to be significant for bit depth if some of its coefficients have absolute value larger than 2. The original SPIHT algorithm repeatedly employs a set partitioning algorithm for identifying and refining significant wavelet coefficients until the rate budget is exhausted. Each successive application of the set partitioning operation decreases the bit depth by one. For each, the set partitioning operation consists of two passes: the sorting pass where the significance of each wavelet coefficient of the image to be encoded is determined with respect to, and the refining pass where the refinement of significant coefficients is performed. To effectively realize these two passes, three lists of information, termed list of significant pixels (LSP), list of insignificant pixels (LIP) and list of insignificant sets (LIS), are maintained at any point of coding. The lists LSP and LIP contain the locations of significant and insignificant wavelet coefficients, respectively. The list LIS contains the root node of the insignificant wavelet tree. The encoding process at each layer in the LSPIHT algorithm also keeps the LSP, LIP, and LIS. Let LSP, LIP, LIS, and be the LSP, LIP, LIS, and threshold value for the encoding of

58 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 7, NO. 1, MARCH 2003 layer, respectively. At the layer 1, the initial LSP is set to be empty, and initial LIP contains,,, and. The initial LIS consists of the wavelet trees having root node in the subbands,, and, respectively. Note that these trees only contain coefficients in the subbands,,,. The threshold value for determining the significance of wavelet coefficients at layer 1 is initialized to be 2, where (2) The SPIHT is then performed over the subbands,,,,, until the rate budget is exhausted. Suppose the encoding of layer is completed, and the encoding of layer is to be done, where. Since the encoding region of layer covers the encoding region of, using the encoding results of layer as initial conditions of layer may enhance the coding efficiency of layer. To accomplish this, we first set the initial be the final. That is, the initial bit depth of layer is identical to the final bit depth of layer. This guarantees the quantization accuracy improves as the LSPIHT algorithm continues. Before applying the SPIHT operation at layer, we initialize LIP, LSP and LIS. The initialization is based on the final LIP, LSP and LIS, and the new subbands (i.e.,, and, ) for the encoding at layer. To perform the initialization, the wavelet trees in the final LIS are first extended to cover coefficients in the new subbands. That is, the wavelet coefficients in the new subbands which are the descendants of leaf nodes of a wavelet tree in the final LIS are included in that wavelet tree. When the coefficients in the new subbands are less than the initial, all the extended wavelet trees remain insignificant. Consequently, we simply initialize LIS as the final LIS. Note that some insignificant coefficients in the new subbands may not also locate in the extended wavelet trees. Those insignificant coefficients and the final LIP are then used to form the initial LIP. Moreover, since no new significant coefficients are created after the wavelet tree extension, we set the initial LS be the final LSP. Note that the extended wavelet trees may not always remain insignificant for any combination of rate and resolution constraints. When some coefficients in the new subbands are larger than the initial, the separation of the corresponding extended wavelet trees into significant coefficients, insignificant coefficients and insignificant wavelet trees are necessary. Consider an extended wavelet tree in which some of wavelet coefficients are larger than the initial. The separation of is also based on the SPIHT. Let LSP, LIP, LIS and be the LSP, LIP, LIS, and threshold for the SPIHT over the extended wavelet, respectively. The procedure of the separation process for then stated as follows. Step 1 Set initial threshold 2, where is given by (2). Set initial LIS, initial LIP be root of, and initial LSP be empty. is Fig. 6. Flowchart of the LSPIHT algorithm. Step 2 Perform the sorting pass to identify the significant coefficients with respect to. Separate the wavelet trees in LIS containing these significant coefficients into significant coefficients, insignificant coefficients and insignificant wavelet trees. Update LIS, LIP and LSP accordingly. Step 3 Perform the refining pass to refine the coefficients in LSP with respect to. Step 4 Reduce by half. If equals initial then stop. Otherwise, go to Step 2. After the separation processes, the LIS is initialized as the set containing the wavelet trees in the final LIS,, and all the unseparated extended wavelet trees. The initial LIP contains the new insignificant coefficients identified after the separation processes, and the insignificant coefficients in the final LIP. Similarly, the initial LSP contains the significant coefficients in the final LSP, and the new significant coefficients found after the separation process. After initial LIS, LIP, and LSP have been determined, the SPIHT is then performed over the subbands,,,,, until the rate budget is exhausted. The same procedure is repeated until the design of highest layer is completed. The complete flowchart of the 2-D LSPIHT algorithm is shown in Fig. 6. The 1-D LSPIHT simply follows the same procedure of the 2-D LSPIHT shown in the figure, except

HWANG et al.: SCALABLE MEDICAL DATA COMPRESSION AND TRANSMISSION 59 TABLE I PERFORMANCE OF THE LIT SYSTEM REALIZED BY LSPIHT FOR VARIOUS TEST MRIS that the 1-D LSPIHT is based on a sequence of 1-D SPIHT [6] operations rather than 2-D ones. Therefore, the design procedure of 1-D LSPIHT is not repeated in this section for the sake of brevity. IV. SIMULATION RESULTS This section presents some numerical results of the LSPIHT technique for the implementation of the LIT systems for telemedicine applications. Table I shows the performance of the transmission system realized by LSPIHT for various MRIs. The dimension of the images is 512 512. Therefore, the full resolution level is. The nine-stage wavelet transform matrices (i.e., ) of the test images are obtained by 9/7-tap filter [1]. The number of layers in the transmission system is. The resolution level of layers 1,2 and 3 are,, and, respectively. Note that since, the reconstructed images at layer 3 are displayed in full resolution. The incremental rates at layers 1, 2, and 3 of the system are,, and, respectively. Therefore, the accumulated rate up to layers 1, 2, and 3 are,, and, respectively. The performance measures for each layer are SNR and rate. Let be the image to be encoded, then the low-pass subband is the desired image to be reconstructed in the layer. Let be the reconstructed image in the layer, and be the mean squared distance between and. Then the SNR in the layer, denoted by SNR, is defined as SNR, where denotes the energy of. Figs. 7 and 8 show the original and reconstructed versions of the MRIs coded by LSPIHT at each layer. Note that the images in the layers 1 and 2 are of lower resolution, and therefore are smaller in size. However, in Figs. 7 and 8, they are zoomed in/up to have same size image as in layer 3 for easy comparison. From Table I and Figs. 7 and 8, it is observed that the LSPIHT technique exhibits high SNR and excellent visual quality in each layer. For comparison purpose, we also implement the SPIHT-based and the JPEG2000-based simulcast systems for the coding of,. In the systems, the low-pass subbands, are encoded independently using the SPIHT and the JPEG2000 techniques, respectively. The encoding process at layer of the SPIHT-based and JPEG2000-based simulcast systems are executed with the rate, where the value of is identical to the incremental rate shown in Table I. The numerical results of the JPEG2000 are obtained from the kakadu implementation with 9/7-tap filter and 16 16 codeblock size. Table II shows the resulting SNR of the reconstructed low-pass subbands,, for various images. It is observed from the table that the JPEG2000-based simulcast system has higher SNR value than that of the SPIHT-based simulcast system at each layer (except only for the MRI01 image at the layer 3, where the SPIHT slightly outperforms the JPEG2000 by 0.6 db). Next we compare the simulcast systems with the LSPIHT-based LIT systems. Note that the rate for the image encoding at layer in the simulcast systems and the LSPIHT-based LIT systems are given by and, respectively. From Tables I and II, we see that in this experiment. Therefore, the comparison is based on the same rate for the image encoding at each layer. We observe from these tables that the JPEG2000-based simulcast system has higher SNR values at layer 1 than those of the LSPIHT-based LIT system. However, the LSPIHT-based LIT system outperforms JPEG2000-based simulcast system at layers 2 and 3. Recall that the encoding process of the LSPIHT-based LIT system at layer 1 is simply the SPIHT operation over. Consequently, the JPEG2000-based simulcast system has superior performance at this layer. Nevertheless, at layers 2 and 3, the LSPIHT algorithm effectively reduces the coding redundancy by taking the encoding results at the previous layers into account. By contrast, the JPEG2000-based simulcast systems may yield large overhead since images at different layers are encoded independently. Because of the overhead, although the JPEG2000 is effective for image coding, the performance of the JPEG2000-based simulcast system is inferior to that of the LSPIHT-based LIT system at layers 2 and 3. To further assess the performance of LSPIHT technique, we also compare the LSPIHT-based scalable system with the simulcast system subject to the same rate for reconstructing each low-pass subband at each layer. Recall in the Fig. 4(b) that, in the LIT systems, the decoders have to accumulate bit streams up to each layer from the base layer in order to reconstruct medical data at that layer. Consequently, the rate for the image reconstruction at layer is given by. On the contrary, the rate for encoding the image at layer in the LIT is only. However, the images at different layers are encoded/decoded independently in the simulcast systems. It then follows from Fig. 4(a) that the rates for encoding and decoding at each layer are identical. The rate for decoding an image at layer in the simulcast systems therefore are also. Consequently, by setting, both the simulcast and the LIT systems have identical rate for image reconstruction. Table III shows the resulting SNRs of the reconstructed of various test images encoded with rate, where the value of,, 2, 3, are given in Table I. Note that the rate for encoding of in the simulcast system is ; whereas, the rate for the encoding of is only in the LSPIHT-based LIT system. From Tables I and III, we find that both the SPIHT-based and JPEG2000-based simulcast systems have slightly higher SNR for each,. Although the simulcast systems perform better in this case, it requires much

60 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 7, NO. 1, MARCH 2003 Fig. 7. Original and reconstructed images of MRI 01 at each layer of LIT shown in Table I. Left column: the original MRI 01 at each layer. Right column: the reconstructed MRI 01 at each layer. higher rate for encoding and transmission. Therefore, they require more storage for saving images and higher bandwidth for transmission. In fact, the total rate for encoding all the images,, in the system is. However, the total rate in the LIT is only. That is, as compared with SPIHT-based and JPEG2000-based simulcast systems, the LSPIHT-based LIT system achieves almost comparable performance with much less rate for image encoding. In addition to MRI encoding, the compression of ECG signals is a major concern in many telemedicine applications. Therefore, in this section, we also apply the LSPIHT for ECG encoding/transmission. Table IV shows the performance of the LIT system realized by LSPIHT for various ECGs. The ECGs are taken from the MIT-BIH arrhythmia database records 100, 117 and 119, sampled at 360 Hz and 11-bit precision. We encode 10 min of data from each of these records. To obtain the wavelet

HWANG et al.: SCALABLE MEDICAL DATA COMPRESSION AND TRANSMISSION 61 Fig. 8. Original and reconstructed images of MRI 02 at each layer of LIT shown in Table I. Left column: the original MRI 02 at each layer. Right column: the reconstructed MRI 02 at each layer. coefficients of ECG, each record is segmented into nonoverlapping and contiguous vectors having identical dimension 1024. Each vector is then independently encoded using the LSPIHT. Therefore, the full resolution level is. The number of layers in the transmission system is. The resolution level of layers 1 and 2 are and, respectively. The performances at each layer are percent root mean square difference (PRD) and CR. For the case of ECG compression, let be the ECG to be encoded, then the low-pass subband is the desired ECG to be reconstructed in the layer. Let be the reconstructed image in the layer, and be the mean squared distance between and. Then the PRD in the layer, denoted by PRD, is defined as PRD (3)

62 IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, VOL. 7, NO. 1, MARCH 2003 TABLE II PERFORMANCE OF THE SIMULCAST SYSTEM FOR VARIOUS TEST MRIS. THE RATE s ;i =1,2,3,FOR ENCODING OF x IS IDENTICAL TO THE r SHOWN IN TABLE I (a) TABLE III PERFORMANCE OF THE SIMULCAST SYSTEM FOR VARIOUS TEST MRIS. THE RATE s ;i =1;2;3FOR ENCODING OF x IS IDENTICAL TO THE r SHOWN IN TABLE I (b) where denotes the energy of. The CR at the layer, denoted by CR, is defined as CR (4) where and are the total number of bits used for representing and, respectively. In this experiment, the CR at layers 1 and 2 are given by 16:1 and 32:3, respectively. Fig. 9 shows the original and reconstructed versions of the ECGs coded by LSPIHT at each layer. The ECGs in the layer 1 are of lower resolution and, therefore, are smaller in size. They are zoomed in/up to have same size image as in layer 2 for easy comparison. From the figure, we observe that the major effect of compression, especially in layer 1, is the smoothing of the background noise. Otherwise, the features of the waveform appear to be faithfully preserved. Table IV also shows the performances of a two-layer simulcast system for encoding ECG signals. Since the JPEG2000 is used only for the image compression, only the performance of the SPIHT-based simulcast system is included in the table. The CR and resolution associated with each layer of the system is identical to those of LSPIHT-based LIT system shown in the table. Consequently, the LIT and simulcast systems have the same performance at layer 1 because both systems employ the SPIHT over the with the same rate at that layer. In addition, it may not be surprising to observe that the LSPIHT-based LIT system has lower PRD value over the SPIHT-based simulcast system at layer 2 based on the same CR. To show the superiority of the LSPIHT algorithm further, we first note that the SPIHT algorithm has been found to outperform various wavelet-based [2], [3] and direct [4] ECG compression techniques. Since those methods [2], [3] are also not both SNR and resolution scalable, (c) Fig. 9. Original and reconstructed ECGs of MIT-BIH 117 at each layer of LIT shown in Table IV. (a) Original ECG. (b) Reconstructed ECG at layer 1. (c) Reconstructed ECG at layer 2. TABLE IV PERFORMANCE OF THE ECG TRANSMISSION SYSTEMS REALIZED BY THE LSPIHT AND SPIHT ALGORITHMS the simulcast systems implemented by those methods may have inferior performance as compared with the LSPIHT-based LIT system. Table V shows the performance of the top layer (layer 2) of the two-layer simulcast systems realized by various ECG encoding methods. The resolutions associated with each layer are identical to those of LSPIHT-based system shown in Table IV.

HWANG et al.: SCALABLE MEDICAL DATA COMPRESSION AND TRANSMISSION 63 TABLE V PRD COMPARISON OF DIFFERENT CODING ALGORITHMS The CR at the first layer is 32:1. From Table V, we find that, with the same CR, the LSPIHT outperforms all the other methods in the table for ECG compression, These numerical results on MRI and ECG demonstrate the effectiveness of the LSPIHT algorithm. [3] M. L. Hilton, Wavelet and wavelet packet compression of electrocardiograms, IEEE Trans. Biomed. Eng., vol. 44, pp. 394 402, 1997. [4] S. M. S. Jalaleddine, C. G. Hutchens, R. D. Strattan, and W. A. Coberly, ECG data compression techniques-a unified approach, IEEE Trans. Biomed. Eng., vol. 37, pp. 329 343, 1990. [5] J. C. Lin, Applying telecommunication technology to health-care delivery, IEEE Eng. Med. Biol. Mag., vol. 18, pp. 28 31, 1999. [6] Z. Lu, D. Y. Kim, and W. A. Pearlman, Wavelet compression of ECG signals by set partitioning in hierarchical trees algorithm, IEEE Trans. Biomed. Eng., vol. 47, pp. 849 856, 2000. [7] A. Said and W. Pearlman, A new, fast, and efficient image codec based on set partitioning in hierarchical trees, IEEE Trans. Circuits Syst. Video Technol., pp. 243 250, June 1996. [8] D. S. Taubman and M. W. Marcellin, JPEG2000 Image Compression Fundamentals, Standards and Practice. Boston, MA: Kluwer, 2002. [9] M. Vetterli and J. Kovacevic, Wavelets and Subband Coding. Englewood Cliffs, NJ: Prentice-Hall, 1995. V. CONCLUSION This paper presents a novel LSPIHT technique for the scalable medical data compression and transmission.the technique allows the resolution and the CR (or the rate) at each layer to be prespecified before the design. The encoding process is then executed based on those specifications. The algorithm is based on a sequence of SPIHT algorithms and, therefore, is simple to implement. Based on the same resolution and CR at each layer, the algorithm outperforms other methods for the compression of MRI and ECG data. Because of its effectiveness and simplicity, the algorithm can be a good alternative for reducing the complexities and costs for realizing the communication network systems for telemedicine applications. REFERENCES [1] M. Antonini, M. Barland, P. Mathieu, and I. Daubechies, Image coding using wavelet transform, IEEE Trans. Image Processing, vol. 1, pp. 205 220, Apr. 1992. [2] A. Djohn, T. Q. Nguyen, and W. J. Tompkins, ECG compression using discrete symmetric wavelet transform, Proc. 17th Int. Conf. IEEE Medicine Biology, pp. 167 168, 1995. Wen-Jyi Hwang (M 94 A 01) received the Diploma degree in electronics engineering from National Taipei Institute of Technology, Taiwan, R.O.C., in 1987, and the M.S. and Ph.D. degrees in electrical and computer engineering from the University of Massachusetts at Amherst, Amherst, in 1990 and 1993, respectively. Since 1993, he has been with the Chung Yuan Christian University, Taiwan, R.O.C., where he is currently a Full Professor with the Electrical Engineering Department. His research interests include signal compression, image processing, and digital communication systems. Ching-Fung Chine received the B.S. degree from the Department of Electronics Engineering, Chung Yuan Christian University, Taiwan, R.O.C., in 1999. She is currently pursuing the Ph.D. degree at the Department of Electrical Engineering in the same university. Her research interests include signal processing, image/video transcoding, image and video technology, digital communication systems, and network applications. Kuo-Jung Li received the M.S. degree form Chung Yuan Christian University, Taiwan, R.O.C., in 2001. His general research interests include the areas of signal processing and wavelet analysis.