Medical images compression: JPEG variations for DICOM standard
1 like6,071 views
The document discusses the DICOM standard for medical imaging, emphasizing its importance for interoperability among medical devices and the various JPEG compression techniques utilized within the standard. It details the different JPEG variations, including lossy, lossless, and near-lossless options, along with their respective algorithms and performance metrics for medical applications. The analysis highlights the trade-offs between compression efficiency, image quality, and processing speed, ultimately noting the advancements in JPEG 2000 for medical imaging needs.
Medical images compression: JPEG variations for DICOM standard
- 1. Medical Images Compression: JPEG variations for DICOM standard Author: Jose Pablo Pinilla Gomez The DICOM (Digital Imaging and Communication in Medicine) is a standard managed by the Medical Imaging & Technology Alliance to ensure the interoperability of medical systems, images, documents and workflow. It was released in 1993 (updated in 2011), to solve the problem with CT (Computed Tomography) or MRI (Magnetic Resonance Imaging) images which cannot be decoded by any means other than those provided by the same manufacturers. This discordance between devices causes unfair competition among medical machines manufacturers by keeping health care facilities from acquiring alternate machines that can lead to device incompatibility. Although there are data “translators” to interconnect different standard systems, their implementation increases the total cost and delay of the system. A standardized system is then necessary for the implementation of a PACS (Picture Archiving and Communication System) that compiles all kinds of acquirable images for medical records. The image file format specified in the current DICOM standard supports RLE (Run Length Encoding), JPEG, JPEG-LS and JPEG-2000 compression and decompression schemes. RLE is a lossless compression scheme that converts repeating byte sequences into two-byte codes with the value of the repeating byte and the negative representation of the number of bytes in the run, this is called a Replica Run; In case a sequence of non-repetitive bytes is found, the resulting code will have one byte with the positive representation of the number of bytes in the sequence followed by the literal sequence (Literal Run)1. RLE can yield good compression ratios for monochrome images but it has a very variable behaviour for RGB and YCbCr images. The use of RLE is often preceded by other algorithms to improve its performance. The Joint Photographic Experts Group created the JPEG standard in 1992. It relies on a five step process to compress pixel data with a configurable parameter of quality percentage. The processing begins by changing the color space from the pixel RGB (Red, Green and Blue) triplet to the YCbCr (Lluminance,Red-Chrominance and Blue-Chrominance) values. As many other image compression schemes, JPEG uses YCbCr color space to keep the lluminance values that are more relevant to the human eye while discarding some of the chrominance information. After decompression, files with “chroma subsampling” are still able to display images with almost unnoticeable changes. A configurable amount of chrominance (hue and saturation) information is discarded by a subsequent downsampling step in JPEG. After subsampling, the next step is to divide the whole image into 8x8 matrices of the same channel (Y, Cb and Cr). Then, every 8x8 pixel block goes through a Discrete Cosine Transform (DCT) computation of coefficients. The DCT gives back 64 values (coefficients) that are related to the “concentration” of one of 64 8x8 blocks with different patterns that can represent the original image when put together, or give a close approximation of it. Each of the predefined patterns is called a “basis function” and it is a two-dimensional set of curves created using the cosine function. The resulting 8x8 matrix with the coefficients is now run through a step of “Quantization”, where every value in the input matrix is divided by a corresponding constant derived from predefined quantization tables and user configuration of quality; the resulting values are rounded to the nearest integers. As stated in the JPEG standard2 “The purpose of quantization is to achieve further compression by representing DCT coefficients with no greater precision than is necessary to achieve the desired 1 DICOM standard Part 5: Data Structures and Encoding. 2011. Annex G. P101. http://medical.nema.org/Dicom/2011/11_05pu.pdf 2 JPEG Still Picture Compression Standard. 1991. p5. http://white.stanford.edu/~brian/psy221/reader/Wallace.JPEG.pdf
- 2. image quality. Stated another way, the goal of this processing step is to discard information which is not visually significant.” The outcome of quantizing the DCT coefficients is that smaller, unimportant coefficients will disappear and larger coefficients will lose unnecessary precision. The final step in JPEG compression is called “Coding” which involves three sub-processes: DC Coding, Zig-Zag Sequence and Entropy Coding. The “DC” coefficient in the matrix, which is the top left-most value, contains a higher integer that is a measure of the average value of the 64 image samples. DC coding replaces every block's DC coefficient with the difference between it and the one of the previous block, this values are often correlated and therefore produce a small difference. Zig- Zag Sequence is a literal zig-zag change in the organization of the matrix, taking advantage of the differentiation between “high-frequency” and “low-frequency” coefficients; the result of this process is that non-zero (low-frequency) and zero (high-frequency) values are grouped so that the following compression steps are more effective. The entropy coding uses the Huffman method which can replace repeating zeroes at the end of a matrix with a single character (byte) and use prefix coding * for other characters according to their frequency of appearance, more frequent characters are coded into shorter values. JPEG's algorithm is very fast, with an average decompression time of 0.9 seconds 3, but it is overall a lossy standard and it reflects it in most of its steps. Color space transformation is an irreversible process because computer calculations of decimal point numbers are always approximations of the real values. Subsampling is a straightforward elimination of data but this step is optional, unlike block splitting in which certain blocks may require “filling” algorithms to complete the 8x8 matrices and thereby altering the decompressed result. Next is the DCT which isn't by itself a lossy process, but forward quantization always is. Quantization is a many-to-one mapping, and therefore is fundamentally lossy. It is the principal source of lossiness in DCT-based encoders.4 Medical Imaging applications often require high precision so that digital alterations can't affect a physician's criteria based on radiology results. This doesn't mean that the compression schemes in Medical PACS have to be lossless, but they need an appropriate level of fidelity. On the other hand, fidelity is a very subjective qualification which is why DICOM also introduced the lossless JPEG scheme as well as the lossless/near-lossless JPEG-LS and the JPEG 2000 later on, as alternatives to the “baseline” JPEG standard5. Lossless JPEG standard was added as a JPEG mode of operation in 1993. It replaces the DCT and other lossy steps (Quantization, Chroma Subsampling) from the original format with a new coding called Differential Pulse Code Modulation (DCPM). DCPM is based on the prediction of an image block based on its surroundings, the values in these predictions are generally accurate which is why the difference between them and the ones achieved in the compression tends to be small. Thereby, this variation of the standard saves these differences and them compresses them even further using the Huffman method. The downside of lossless JPEG is that it cannot achieve compressions ratios comparable to those of baseline JPEG compression. * Prefix property states that there is no valid code word in the system that is a prefix (start) of any other valid code word in the set. 3 JPEG 2000 still image coding versus other standards. 2000. http://www.jpeg.org/public/wg1n1816.pdf 4 SCHROEDER, Mark D.. JPEG Compression Algorithm and Associated Data Structures. 1997. http://akbar.marlboro.edu/~mahoney/courses/Fall01/computation/compression/jpeg/jpeg.html 5 The JPEG Committee. Medical Imaging. 2007. http://www.jpeg.org/apps/medical.html
- 3. In 1999 the Joint Photographic Experts Group released the JPEG-LS standard as a solution to achieve better lossless compression. It uses the prediction, RLE, and context modelling (surrounding blocks dependency) coding techniques included in the LOCO-I algorithm created by Hewlett Packard (HP). This scheme helps to achieve a better compression while keeping the integrity of information in accord to user's configuration6. Therefore, this standard is classified as lossless/near-lossless, where near-lossless refers to a change from 1 to 5 between the original RGB values and the uncompressed versions7. It takes approximately the same time to decompress a JPEG-LS file and a lossless JPEG but the compression ratio achieved is better. The latest standard included in the DICOM compatibility was JPEG 2000 which in addition to having a lossless mode, it includes a set of features that are advantageous in Medical Imaging applications, such as: Regions of Interest, scalability(Low-resolution previewing) and extensive MetaData. The differences between baseline JPEG and JPEG 2000 algorithms reflect a new approach in which the objective of higher compression is always accompanied by a lossless alternative. That is, the Color Space Transformation can either be irreversible or reversible by removing the quantization errors by using integers; instead of block splitting, JPEG 2000 uses variable size “tiles” to remove undesired block filling; the DCT is replaced by a Wavelet Transform (WT) with lossy and lossless variations, again by removing quantization; and it also runs a coding algorithm afterwards. The WT differs from the DCT, in that, instead of cosine functions it uses another kind of oscillating functions called wavelets. A wavelet's amplitude starts from zero, reaches its maximum value and then fades back to zero creating a “frequency pulse”. The effect of this change in the compression algorithm is that this transform's coefficients are centred around zero (can be easily coded), with a very few large coefficients. Nevertheless, JPEG 2000's algorithms make image compression and decompression relatively slow, taking an average of 4.3s to decompress an image. Although there is a limitation to JPEG file formats in the DICOM standard, there is a high range of possibilities that rise from the different JPEG versions and the quality configuration capabilities; in addition, JPEG is now working in collaboration with DICOM to further improve this integration so that Medical Imaging necessities are taken under consideration8. But still, when it comes to image quality there is a lot of subjectivity in order to define how lossy an image can be, taking into account that lossless algorithms cannot yield as good compression results as those with lossy schemes. Therefore, the decision to implement a lossy or lossless scheme, and how lossy it can be is now turning into a decision of what features and performance is required by the application. Lossless and near-lossless standards with good speed, compression ratio, and new features have the Medical Imaging industry as one of its targets. However, in order to improve any of those features the system has to sacrifice performance in another, for example, the average compression ratio for the lossless configurations of JPEG(2.09), JPEG-LS(2.98) and JPEG 2000(2.50) differ, making JPEG-LS the best choice for lossless compression, also due to it's fast performance. But if a certain amount of quality can be compromised (near-lossless), the JPEG algorithm will give better results in considerably less time. Although, when compared to JPEG 2000 that offers the “Medical-Imaging-oriented” features, JPEG 2000 will outperform JPEG by giving out the same compression ratios with more quality settings, taking into account that this superiority is toned down by the amount of time that it takes to decompress. 6 The JPEG Committee. Lossless JPEG. JPEG-LS. 2007. http://www.jpeg.org/jpeg/jpegls.html 7 Lossless, Near-lossless and Lossy Compression of PTM Images. HP. 2001. http://citeseerx.ist.psu.edu/viewdoc/download?doi=10.1.1.22.1635&rep=rep1&type=pdf 8 The JPEG Committee. Medical Imaging. 2007. http://www.jpeg.org/apps/medical.html