Block-Wise Context-Adaptive Re-Indexing for Lossless Compression of Grayscale Medical Images

Many image compression algorithms have been developed by computer scientists depending on fundamental properties of images such as spatial correlation and statistical redundancy. In case of algorithms that relay on spatial correlation, they focus on neighboring pixels whose values are often similar or related like gradual color changes in images of Walls or skins [10].

On the other hand, when pixel value or patters are repeated across over an Image this feature well known as statistical redundancy, which is made a Solid base for a compression Algorithm, as Like sky Image which it has many pixels have same shade of blue [11].

2.1 Spatial correlation image compression approaches

Here are a few image compression approaches that utilized spatial correlation.

2.1.1 JPEG

It is a classic image compression technique; it is a standard found that was found by the Joint Photographic Exports Group, which classified as lossy compression technique, which removes unnecessary visual information to reduce image file size. The first step of the compression process is to transform the image color space from Red, Green, Blue (RGB) to YCbCr in order to split the brightness channel (Y) from the color channels (Cb, Cr) which enables the process to reduce color data (Chroma subsampling) that is less perceptible to the human eye. After that, the process involves Discrete Cosine Transform that change spatial domain into frequency domain components [12]. Discrete Cosine Transform (DCT) takes action by splitting the image into non-overlap blocks each of which is 8 by 8 pixels in size, the transformation focuses the signal energy into few significant coefficients to make it easier to discard less important data during compression [13], and the next important step is quantization. This is where the less important high-frequency coefficient is split up by the quantization matrix and bounded off to zero. This step is where most of the "loss" comes from and allows for a significant reduction in data. Finally, a lossless entropy method used to compress the remaining coefficients such as Huffman and arithmetic coding to make the image file size even smaller [14, 15].

2.1.2 Wavelet-based compression

This technique is considered a sophisticated image compression technique where the image is broken down in to different frequency components by employing a wavelet transform. It is well implemented by the JPEG 2000 standard.

The first step starts by processing the entire image at once in order to break down the image into groups of wavelet coefficients representing the high-frequency coefficient (fine lines, edges, and textures) and the low-frequency coefficient (the overall color and shape).

Secondly, the quantization step is where the actual compression happens at this lossy step, where the quantization sets the high-frequency coefficients to zero or their precision significantly decreased, while the Low-frequency coefficients are maintained to keep the image's essential features [16]. The final step is to make the image file even smaller by using a lossless compression method such as arithmetic coding to encode the quantized coefficient (where many of them are zero now) [17].

2.2 Statistical redundancy image compression approaches

Here are several methods for image compression that make use of statistical redundancy.

2.2.1 Huffman coding

David Huffman created this compression algorithm. In 1952, LossLess compression algorithm was characterized by efficiency and simplicity. It uses for compress images and multimedia widely, where it gives the more frequent symbols shorter binary codes and the longer codes are given to less frequent symbols [18]. Many modern studies continue to improve the Huffman algorithm by integrating this algorithm with different techniques like adaptive coding and hybrid models. Singh et al proposed an enhanced adaptive Huffman coding algorithm to improve compression efficiency in real-time applications [19].

2.2.2 Run length encoding

RLE it is considered a simple lossless compression technique that store a sequence of identical values with a single value and its duplicates. In the context of image compression, it uses a pair value to represent consecutive pixels of the same color [20]. This technique is suitable for images that have large areas of solid, uniform color that contain consecutive pixels or pixel values that are identical with small variation, such as black and white images, icons, and line drawings [21].

The effectiveness of the proposed algorithm is evaluated by applying it to several different medical images (see Figure 3).

Figure 3. Dataset medical images

As a lossless compression algorithm, the original image is identical to the decoded compressed image, and no degradation evaluation is required. A compression ratio was adopted as a performance metric, the compression ratio is defined as the ratio of the original image size to the compressed image size. For that, the only performance metric used to evaluate the algorithm is the compression ratio. The performance evaluation process also takes into account the use of different pixel block sizes (8 × 8 and 4 × 4); see the Tables 2 and 3.

Table 2. Compression ratio (4 × 4) pixel block

Table 3. Compression ratio (8 × 8) pixel block

Tables 2 and 3 reveals information about the original image size and compressed image size utilizing two pixels block sizes (8 × 8 and 4 × 4); these information can be summarized as follows.

1- High compression efficiency with 4 × 4 blocks

Focusing on the compression ratio, the (4 × 4) block consistently outperforms the (8 × 8) block, which gives an indication that the (4 × 4) block provides stronger compression, and this algorithm is pixel block size sensitive. When converting from an (8 × 8) to a (4 × 4) block, the compression ratio commonly improves by (0.7 to 1.1).

2- Most compressible and least compressible images

The most compressible image in the dataset is the “knee Magnetic Resonance Imaging (MRI) 2” image which has the highest compression ratio for both block sizes (2.757 for 8 × 8 and 3.7434 for 4 × 4).

While the “Cervical Spine” image has the lowest compression ratio (1.4037 for 8 × 8 and 2.068 for 4 × 4) suggesting that the image has more complex structures and is harder to compress effectively.

Table 4. Compression ratio at (8 × 8) and (4 × 4) pixel blocks and Huffman compression technique

Table 4 shows the values of the compression ratio for the proposed algorithm (using two-block sizes 4 × 4 and 8 × 8) and the Huffman compression technique. There are a number of distinct trends to note.

For all the images of the dataset, the compression ratio produced by the Huffman technique is significantly lower than the compression ratio produced by the proposed algorithm, whether using 8 × 8 or 4 × 4 blocks. That indicates that the proposed technique achieves more efficient size reduction than standard Huffman coding.

4.1 Complexity and overhead analysis

As the suggested algorithm test its performance in block sizes of (4 × 4) and (8 × 8), and the (4 × 4) block configuration requires four times of header as (8 × 8) headers needed, that leads to higher header overhead as a result of the number of blocks per image compared to (8 × 8) configuration. Furthermore, as the suggested re-indexing approach depend on simple frequency sorting instead of complex iterative transforms; it is classified as computationally efficient. The confirming lossless recovery was verified for all test images using the MD5 (pixel-by-pixel difference) between the original and the decompression images.

4.2 Comparing with JPEG and JPEG2000

For expanding the performance evaluation of the suggested algorithm, the comparison was made with the JPEG and JPEG2000 standards (as shown in Table 5). For ensuring statistical consistency, these baselines were create using the same set of medical images. The number in Table 5 shows a clear advantage of the proposed block-based re-indexing in most cases overcoming the conventional frequency – based transforms, especially in capturing localized statistics of MRI scans.

Table 5. Comparison with JPEG and JPEG2000

4.3 Evaluate the proposed algorithm on standard clinical Digital Imaging and Communication in Medicine medical modalities

The proposed algorithm has been subject to another examination to further validate the empirical resilience compressing ten images from the Digital Imaging and Communication in Medicine (DICOM) dataset using the following evaluation metrics: Compression Ratio (C.R.), Bits Per Pixel (bpp), Mean Squared Error (MSE), Peak Signal to Noise Ratio (PSNR), and Structural Similarity Index (SSIM).

Below Figure 4 presents a set of medical images taken form DICOM database for evaluate the proposed algorithm and Tables 5 and 6 demonstrate the test result using both of (4 × 4) and (8 × 8) block sizes.

Figure 4. A set of Digital Imaging and Communication in Medicine (DICOM) dataset medical images

The experimental data show the block size is an architectural trade-off inherently. That it relates to compression performance where it shows stability and Tables 6 and 7 robustness at a block size of (8 × 8) where the CR has range of (2.07 – 2.20) and improves in bit-rate value (3.63 to 3.87) bpp across most of the clinical images. On the other hand, when swapping to a block size (4 × 4) an improvement in the CR values was notice. Where its range is 2.80 to 3.12, and the bit rate show a promising value where it's getting down to range of 2.57 to 2.85 bpp.

The above algorithm facts are highly justified as the local statistical correlation are vastly maximized within the pixel neighborhoods for smaller block dimensions. Leading to minimizing the number of unique intensity values, which dynamically optimizes the header representation and permits tighter re-indexing. The most important thing to be highlighted throughout all the recorded values is that the evaluation metrics achieved MSE = 0, SSIM = 1 ad an infinite value for PSNR. These values demonstrate the algorithm lossless data integrity of the context-aware mechanism, which is essential for accurate medical diagnostics.

Table 6. Evaluation metrics of Digital Imaging and Communication in Medicine (DICOM) image data base at block size (4 × 4)

Table 7. Evaluation metrics of Digital Imaging and Communication in Medicine (DICOM) image data base at block size (8 × 8)

4.4 Clinical workflow and real-world deployment (Picture Archiving, and Communication Systems and telemedicine)

Modern healthcare information systems, including Hospital Information System (HIS), Picture Archiving, and Communication Systems (PACS), where their neutral work is to manage and create a huge number of medical imaging data such as DICOM files, to handle such a task they rely on two essential requirements: high transmission bandwidth and large storage capacity. The integration of the suggested algorithm into these medical imaging information systems in smoothly and continuously way. Algorithm structure depend on the Global Occurrence Table with optimized header design in the course of telemedicine data transmission, the communication overhead is reduced by the use of a compressed bitstream as a limited bandwidth existed commonly. Additionally, the enabling of rapid, localized, chunk-based analyzing depends on small block structures like (4 × 4) and (8 × 8). Within modern healthcare information systems, these capabilities play vital role in supporting efficient real-time clinical workflows. Allowing for quick random access to particular regions of interest (ROI) and progressive image rendering without need to decompress the whole medical volume. Consequently, a lossless reconstruction of the compressed image is guaranteed by low-complexity header deployment while essential diagnostic details are maintained for accurate medical decision.

4.5 Statistical and mechanism analysis of block size performance

The analyzing the balance between header overhead and local entropy is considered essential to understand the reason for why the 4 × 4 block overcomes the 8 × 8 block. Grayscale medical image characterized (from a mathematical view) with high spatial correlation within ultra-localized neighborhoods. To obtain (per block) fewer unique values and local frequency concentration, it must minimize local pixel variance by reducing block size to 4 × 4. This effectively lowers the required bit depth for pixel values. Despite choosing a 4x4 block size, which it increases the number of blocks four times that of 8 × 8, which increases total header overhead. This cost can be compensated by the massive bit saving achieved by encoding the highly compressed local pixel arrays. It is also worth noting that this approach can be applied for homogeneous medical textures in a perfect way, it faces a difficulty in highly chaotic natural images. Consequently, and this local-rank method is specifically tailored for medical dataset.

The authors would like to express their gratitude to editor and reviewers, as they provide great guidance to improve this manuscript.