A Differential Compression Method and System for Medical Images Driven by Structural Importance Maps
By employing a differentiated compression method based on structural importance maps, the problems of loss of key diagnostic information and coding block effect under high compression ratios in existing technologies are solved, achieving efficient, interpretable, and auditable compression of medical images.
Patent Information
- Application Number
- CN202611096612.4
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2026-07-23
- Publication Date
- 2026-11-10
AI Technical Summary
Existing medical image compression technologies cannot ensure the preservation of critical diagnostic information at high compression ratios, and existing methods suffer from uncontrollable image loss or coding block effects, failing to effectively protect diagnostic structural information in medical images.
By employing a differential compression method based on structural importance maps, unsupervised geometric structure analysis is used to generate structural importance maps. Combined with mapping functions and adaptive smoothing, differential coding quantization step size offsets are generated to ensure lossless compression of key structures and eliminate coding block effects.
It achieves absolutely lossless compression of critical diagnostic structures at high compression ratios, while eliminating coding block effects, improving the interpretability and auditability of the compression process, and is applicable to a variety of medical imaging modalities.
Smart Images

Figure CN123075178A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of medical image processing and compression coding technology, and in particular to a medical image differential compression method and system driven by structural importance map. Background Technology
[0002] With the continuous improvement of resolution in medical imaging equipment such as CT and MRI, the storage and transmission of massive amounts of image data places extremely high demands on image compression technology. According to review articles such as "Application and Progress of Image Compression Algorithms in Medical Imaging," professional institutions such as the Royal College of Radiology and the Röntgen Society have issued clinical recommendations for the use of lossy image compression, indicating that the limited introduction of efficient lossy compression, while ensuring diagnostic quality, is clinically feasible. This provides an important application background for this invention: the medical community's acceptance of lossy compression is conditional—it must ensure that critical diagnostic information is not lost.
[0003] Currently, medical image compression methods encompass multiple technical directions, including global compression, region of interest (ROI)-based compression, and deep learning-based compression. Global compression, aiming for overall pixel fidelity, treats all image content indiscriminately, failing to prioritize the protection of critical diagnostic regions at high compression ratios. ROI-based compression segmentes the image into ROIs and non-ROIs, employing different compression strategies for each. However, review literature clearly points out that ROI segmentation must be closely integrated with lesion regions, as the segmentation method directly impacts diagnostic value. Furthermore, deep learning-based segmentation algorithms suffer from poor interpretability, placing an additional burden on ROI methods in clinical deployment due to the need for algorithm selection and evaluation. While end-to-end deep learning-based compression methods can achieve high compression ratios, review literature also indicates that these methods struggle to fully control the compression process, requiring additional supervision to avoid uncontrollable image loss. This inherent limitation restricts their application in medical diagnostic scenarios requiring strict traceability.
[0004] It is worth noting that existing technologies employ a class of image compression methods that utilize structural analysis. These methods typically extract sparse structural representations first, then discard most of the original pixels, and reconstruct the image using a mathematical model at the decoding end. In contrast, this invention retains all original pixels and allocates the bitrate by differentially controlling the quantization step size. This architecture allows the invention to be directly embedded into existing PACS and DICOM standard workflows. Furthermore, before the compression process occurs, this invention can explicitly define and decouple "diagnostic structural information" and "visual redundancy information" in medical images in an interpretable and auditable manner, thereby achieving absolute lossless compression of the former and extreme compression of the latter during compression.
[0005] In addition, existing technologies utilize structural tensors to assist in image compression decisions. For example, the variance of the structural tensor matrix elements within an image patch is used to assess the severity of color changes, thereby dividing the image patch into lossless and lossy compression regions. However, such methods have inherent flaws: First, variance, as a scalar statistic, loses the directional information in the structural tensor and cannot distinguish between structural types with different diagnostic values, such as "edges," "corners," and "flat areas." In medical images such as CT and MRI, there are key diagnostic structures with gradual gray-level changes but clear directionality (such as the blurred edges of microcysts or the boundaries of soft tissues with gradual density changes). The gradual gray-level changes of these structures result in low variance values, but their directional structural features are still significant. Variance methods systematically omit these structures, causing them to be incorrectly classified into lossy compression regions, resulting in irreversible loss of diagnostic information. Second, its hard binary classification mechanism causes abrupt changes in quantization parameters at the boundaries of different regions, which easily leads to coding block effects and lacks smooth transition processing for boundary regions.
[0006] For example, at extremely high compression ratios, crucial diagnostic information such as minute spiculations at the edges of lung nodules and fine calcifications at vascular bifurcation points are easily smoothed out along with flat background areas. Furthermore, methods that segment and compress based on pixel grayscale similarity may incorrectly classify lesions as background areas due to their similar grayscale to surrounding tissues, resulting in irreversible loss of diagnostic information.
[0007] Therefore, a completely new compression control method is urgently needed to resolve the above contradictions. Summary of the Invention
[0008] In view of this, the purpose of the present invention is to provide a medical image differential compression method and system based on structural importance map-driven approach, which aims to solve the problem in the prior art that the medical image compression process cannot achieve a high compression ratio while ensuring that the anatomical structural information with high diagnostic value in the image is absolutely lossless in an interpretable and auditable manner.
[0009] To achieve the above objectives, the present invention adopts the following technical solution: a medical image differential compression method driven by structural importance maps, comprising the following steps:
[0010] S1 receives the raw medical image data to be compressed;
[0011] S2, perform unsupervised geometric structure analysis on the original medical image data to generate a spatial point-by-point structural importance map that reflects the complexity of anatomical structures; wherein, the geometric structure analysis derives scalar values that reflect the type of local geometric structures by calculating the eigenvalue relationships of local structural tensors;
[0012] S3, based on the structural importance map, generate differentiated coding quantization step size offsets for different regions in the image through a preset mapping function;
[0013] S4, perform structure-adaptive smoothing on the quantization step offset, using an anisotropic filter guided by the original image structure tensor, allowing smoothing in the tangent direction of the structure edge to eliminate the coding block effect at the region boundary, and suppressing smoothing in the normal direction to maintain the importance boundary.
[0014] S5, the processed quantization step offset is output to the downstream encoder to drive it to perform differential compression encoding on the original medical image data.
[0015] In a preferred embodiment, the geometric structure analysis is performed by calculating the structure tensor and its eigenvalues of the local neighborhood of each pixel. The relationship between the two structures is used to distinguish different structural types such as edges, corners, and flat areas, and to assign different importance values accordingly.
[0016] In a preferred embodiment, the generation of the structural importance map adopts a multi-scale fusion strategy: a multi-scale pyramid is constructed on the original image, structural importance maps are calculated at each scale, and then merged into the final map through a preset fusion strategy.
[0017] In a preferred embodiment, the mapping function is a piecewise linear function that divides the importance value into multiple intervals, each interval corresponding to a different quantization step size offset, and the interval with higher importance tends to have a more negative quantization offset.
[0018] In a preferred embodiment, when the mapping function converts the importance metric into a quantization step offset, it integrates two factors, the structural importance metric and the local texture complexity, to make a fusion decision. For regions with the same importance level but different texture complexities, differentiated quantization step offsets are generated. For regions with the same importance level, the higher the texture complexity, the larger the quantization offset, and the smoother the texture, the smaller the quantization offset.
[0019] In a preferred embodiment, before generating the quantization step offset, a regional structural consistency detection is performed on the structural importance map: for isolated pixels marked as high importance, if the density of high importance pixels in their neighborhood is lower than a preset threshold, their importance is forcibly downgraded to be consistent with the surrounding background area.
[0020] In a preferred embodiment, in the structure-adaptive smoothing process, the diffusion tensor of the anisotropic filter is determined by the structure tensor of each pixel in the original image, and the diffusion coefficient has different values in the edge tangent direction and the normal direction.
[0021] In a preferred embodiment, before generating the structural importance map, a modality adaptation step is included: a lightweight image classification network automatically identifies the imaging modality type of the input medical image and automatically adapts the hyperparameters of the structural analysis algorithm based on the identification results.
[0022] This invention also provides a medical image differential compression system driven by structural importance maps, characterized in that running the aforementioned medical image differential compression method driven by structural importance maps includes:
[0023] A structural analysis module is used to perform unsupervised geometric structural analysis on raw medical images and generate structural importance maps;
[0024] A parameter decision module is used to generate differentiated coded control parameters based on the structural importance map and a preset mapping function;
[0025] A structure adaptive smoothing module is used to perform anisotropic filtering smoothing processing on the encoding control parameters based on the original image structure tensor.
[0026] An encoding driver interface is used to output the processed encoding control parameters to the downstream encoder.
[0027] The present invention also provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the aforementioned method for differential compression of medical images based on structural importance maps.
[0028] Compared with the prior art, the present invention has the following beneficial effects:
[0029] 1. It reconciles the contradiction between lossless diagnosis of key structures and extreme background compression, while eliminating coding block effects at region boundaries through structural adaptive smoothing, suppressing noise interference through isolated point removal, and achieving finer bitrate allocation through texture complexity correction.
[0030] 2. By analyzing the eigenvalue relationships of the structural tensor, we can capture key diagnostic structures with smooth gray-level changes but clear directionality. Existing variance-based methods systematically miss such structures due to the smooth gray-level changes.
[0031] 3. The structural importance map is automatically generated through unsupervised geometric structure analysis, without the need for manual annotation or reliance on specific lesion detection models. This avoids the additional burden of segmentation algorithm selection and segmentation result evaluation in traditional ROI methods, while also possessing complete interpretability.
[0032] 4. The atlas generation is completely unsupervised, independent of any specific diagnostic task or lesion annotation, and universally applicable to different modalities and anatomical locations. Furthermore, as it is itself an intuitive explanation of compression strategies, it meets the fundamental need for trust traceability in the medical field.
[0033] 5. As an encoding control module, the solution can be used with various existing or future encoders, seamlessly integrating into the existing hospital imaging system workflow. Attached Figure Description
[0034] Figure 1 This is an overall flowchart of a preferred embodiment of the present invention. Detailed Implementation
[0035] The present invention will be further described below with reference to the accompanying drawings and embodiments.
[0036] It should be noted that the following detailed descriptions are illustrative and intended to provide further explanation of this application. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application pertains.
[0037] It should be noted that the terminology used herein is for the purpose of describing particular implementations only and is not intended to limit the exemplary implementations according to this application; as used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise; furthermore, it should be understood that when the terms “comprising” and / or “including” are used in this specification, they indicate the presence of features, steps, operations, devices, components and / or combinations thereof.
[0038] Example 1: Basic Implementation Based on Structure Tensor
[0039] refer to Figure 1 This embodiment runs on the compression server of a hospital PACS system. The server is equipped with an HEVC hardware encoder that supports CTU-level quantization parameter offset interface. The original CT images are obtained from the image acquisition device or PACS storage in DICOM format.
[0040] Taking a chest CT plain scan image as an example, the downstream encoder uses the HEVC Intra encoder.
[0041] S1: Receive image
[0042] Receive one 16-bit grayscale DICOM format CT image of pixels.
[0043] S2: Generate a structural importance map
[0044] For each pixel of the image, in one The structure tensor is calculated within a neighborhood window. Eigenvalue decomposition is performed on the tensor to obtain... and Using the formula Calculate the structural importance scalar value for each pixel. ,in To prevent division by zero for extremely small positive values, apply the following to all pixels. The values are normalized to the [0,1] interval to form a structural importance map. In this map, high-value regions such as blood vessel bifurcation and tissue boundaries are close to 1, while low-value regions such as uniform background are close to 0.
[0045] S3: Isolated Node Removal Processing
[0046] The generated structural importance map is subjected to region consistency detection. Specifically, for each The pixels, check its Within the field The number of pixels. If the number is lower than a preset threshold (e.g., 3), the pixel is determined to be a false high-importance point caused by noise, and is removed. The value is forcibly set to 0. After this processing, false high-importance pixels generated in a uniform background due to factors such as CT quantum noise are effectively eliminated, avoiding abrupt fluctuations in quantization parameters in the background area.
[0047] S4: Generate differential coding control parameters
[0048] Use a piecewise linear mapping function ,Will Value and local texture complexity Commonly mapped to quantization step size offset Among them, local texture complexity For this pixel Local variance of grayscale values within the domain.
[0049] The basic mapping rule is: when (Extremely important area) =-6 (approaching lossless); when 0. (Highly Important Area) ;when (Medium-important area) (Maintain the baseline); when (Low importance area / background) (Tend towards high compression).
[0050] Based on this, texture complexity correction is introduced: ,in To correct the strength coefficient (e.g., take 2). This normalizes to a local texture complexity of [0,1]. This correction results in a slightly larger quantization step size for textured complex regions and a slightly smaller quantization step size for textured smooth regions within the same importance level, achieving a finer bitrate allocation.
[0051] Will The graph is aggregated and averaged in units of code blocks (CTUs) to generate the final coding control parameters.
[0052] After this correction, smooth edge regions such as blood vessel walls achieve a lower quantization step size (higher fidelity), while high-texture regions such as bone trabeculae have an appropriately increased quantization step size while maintaining acceptable diagnostic quality, thus achieving a more refined bitrate allocation.
[0053] S5: Adaptive Smoothing Processing
[0054] An anisotropic diffusion filter guided by the local structure tensor of the original CT image is used to filter the above. The image is smoothed. The diffusion tensor of this filter is determined by the structure tensor of each pixel in the original image: in the edge tangent direction, the diffusion coefficient is set to a larger value (e.g., 1.0) to allow for a smooth transition of quantization parameters and eliminate blockiness; in the edge normal direction, the diffusion coefficient is set to a smaller value (e.g., 0.1) to maintain the clarity of importance boundaries. After this processing, The transition between regions of different importance becomes natural, effectively avoiding the block effect that may occur in the decoded image, while the importance boundaries of each region remain clear.
[0055] S6: Drive encoder
[0056] The CTU after the above processing The parameters are written to the CTU-level quantization parameter offset interface of the HEVC encoder to drive it to complete differential compression encoding.
[0057] In the above process, steps two through five generate a differential coding control strategy for the original CT image. (Atlas). In step six, the HEVC encoder performs actual compression coding on the original CT image according to this strategy: in high-importance areas (such as vascular bifurcation, nodule edges). When the value is negative, the quantization step size is extremely small, and the encoded bitstream retains the complete diagnostic information of the region; in low importance regions (such as uniform background). With a positive value and a large quantization step size, the encoded bitstream significantly reduces the storage footprint of this region. The final compressed bitstream has a much smaller total data volume than the original image, and the information fidelity of the diagnostic key regions is deterministically guaranteed.
[0058] Example 2: Preferred Implementation of Introducing Multi-Scale Fusion
[0059] To more robustly capture anatomical information at different levels, this embodiment introduces a multi-scale fusion mechanism based on step 2 of embodiment 1.
[0060] When generating the structural importance map, a three-layer Gaussian pyramid is first constructed on the original image. A structural importance map is then calculated for each layer. Next, the maps at each scale are upsampled to their original size and merged using a fusion strategy that takes the maximum value. This strategy ensures that structures identified as important at any scale (such as organ outlines at a coarse scale and tiny blood vessels at a fine scale) are preserved in the final map. Subsequent steps are the same as in Example 1.
[0061] Example 3: Preferred Implementation of Introducing a Cross-Modal Adaptive Mechanism
[0062] To improve the versatility and adaptability of the method of the present invention across different medical imaging modalities (such as CT, MRI, and ultrasound), this embodiment adds modality adaptive processing before step two of embodiment one.
[0063] First, the input image is scaled to a uniform size and fed into a pre-trained lightweight image classification network (such as MobileNetV3-Small). This network outputs the probability distribution of the modality to which the image belongs (CT, MRI-T1, MRI-T2, ultrasound, etc.). Then, based on the identified modality type, the network automatically selects an appropriate structural tensor from a pre-defined parameter library to calculate hyperparameters. For example, CT images have sharp anatomical structure edges, so a smaller neighborhood window (e.g., 3×3) is used; ultrasound images have significant speckle noise, so a larger neighborhood window (e.g., 7×7) is used, and Gaussian pre-smoothing is enabled to enhance the robustness of structural analysis. Subsequent steps are the same as in Example 1.
Claims
1. A medical image differential compression method driven by structural importance maps, characterized in that, Includes the following steps: S1 receives the raw medical image data to be compressed; S2, perform unsupervised geometric structure analysis on the original medical image data to generate a spatial point-by-point structural importance map that reflects the complexity of anatomical structures; wherein, the geometric structure analysis derives scalar values that reflect the type of local geometric structure by calculating the eigenvalue relationships of local structural tensors; S3, based on the structural importance map, generate differentiated coding quantization step size offsets for different regions in the image through a preset mapping function; S4, perform structure-adaptive smoothing on the quantization step offset, using an anisotropic filter guided by the original image structure tensor, allowing smoothing in the tangent direction of the structure edge to eliminate the coding block effect at the region boundary, and suppressing smoothing in the normal direction to maintain the importance boundary. S5, the processed quantization step offset is output to the downstream encoder to drive it to perform differential compression encoding on the original medical image data.
2. The medical image differential compression method based on structural importance map driven by claim 1, characterized in that, The geometric structure analysis described above calculates the structure tensor and its eigenvalues of the local neighborhood of each pixel. The relationship between the two structures is used to distinguish different structural types such as edges, corners, and flat areas, and to assign different importance values accordingly.
3. The medical image differential compression method based on structural importance map driven according to claim 1, characterized in that, The generation of the structural importance map adopts a multi-scale fusion strategy: a multi-scale pyramid is constructed on the original image, structural importance maps are calculated at each scale, and then the maps are merged into the final map through a preset fusion strategy.
4. The medical image differential compression method based on structural importance map driven according to claim 1, characterized in that, The mapping function is a piecewise linear function that divides the importance value into multiple intervals. Each interval corresponds to a different quantization step offset, and the higher the importance of the interval, the more negative the quantization offset tends to be.
5. The medical image differential compression method based on structural importance map driven according to claim 4, characterized in that, When converting importance metrics into quantization step offsets, the mapping function integrates two factors—structural importance metrics and local texture complexity—to make a fusion decision. For regions with the same importance level but different texture complexities, differentiated quantization step offsets are generated. For regions with the same importance level, the higher the texture complexity, the larger the quantization offset, and the smoother the texture, the smaller the quantization offset.
6. The medical image differential compression method based on structural importance map driven according to claim 1, characterized in that, Before generating the quantization step offset, the structural importance map is first subjected to regional structural consistency detection: for isolated pixels marked as high importance, if the density of high importance pixels in its neighborhood is lower than a preset threshold, its importance is forcibly downgraded to be consistent with the surrounding background area.
7. The medical image differential compression method based on structural importance map driven according to claim 1, characterized in that, In structure-adaptive smoothing, the diffusion tensor of the anisotropic filter is determined by the structure tensor of each pixel in the original image, and the diffusion coefficient has different values in the edge tangent direction and the normal direction.
8. The medical image differential compression method based on structural importance map driven according to claim 1, characterized in that, Before generating the structure importance map, a modality adaptation step is included: a lightweight image classification network automatically identifies the imaging modality type of the input medical image and automatically adapts the hyperparameters of the structure analysis algorithm based on the identification results.
9. A medical image differential compression system driven by structural importance maps, characterized in that, Running any one of the structural importance map-driven differential compression methods for medical images according to claims 1 to 8, comprising: A structural analysis module is used to perform unsupervised geometric structural analysis on raw medical images and generate structural importance maps; A parameter decision module is used to generate differentiated coded control parameters based on the structural importance map and a preset mapping function; A structure adaptive smoothing module is used to perform anisotropic filtering smoothing processing on the encoding control parameters based on the original image structure tensor. An encoding driver interface is used to output the processed encoding control parameters to the downstream encoder.
10. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements a medical image differential compression method based on structural importance map driven by any one of claims 1 to 8.
