CN-122027806-A - Full-field pathological image self-adaptive differential compression method based on AI semantic importance and application thereof

CN122027806ACN 122027806 ACN122027806 ACN 122027806ACN-122027806-A

Abstract

The invention relates to the technical field of image compression and artificial intelligence, and discloses a full-field pathological image self-adaptive differential compression method based on AI semantic importance and application thereof. Aiming at the problems of high storage cost and loading katon caused by huge full-view pathological image files, the method acquires a low-resolution thumbnail of the full-view pathological image files, inputs a pre-trained semantic evaluation model, generates a two-dimensional probability mask comprising three importance levels of extremely low, medium and high, maps the coordinates of the two-dimensional probability mask to the high-resolution image and divides the high-resolution image into grid blocks, extracts color means of the extremely low-importance blocks for bypass replacement, respectively performs lossy and lossless compression on the medium and high-importance blocks, generates a smooth transition area at the junction, and finally generates a compressed file embedded with an index dictionary. The invention effectively reduces the storage occupation and transmission delay of pathological data and ensures the diagnosis fidelity of the core focus area.

Inventors

HUANG QIANG
WANG ZIHAN
CHEN YULING
JIN JIE
KUANG GUOTAO

Assignees

深圳市生强科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260415

Claims (10)

1. The full-field pathological image self-adaptive differential compression method based on the AI semantic importance is characterized by comprising the following steps of: Acquiring a low-resolution level image of the full-view pathology image as a thumbnail; Inputting the thumbnail into a pre-trained semantic evaluation model, and outputting semantic importance scores of all areas; Generating a two-dimensional probability mask comprising three levels of extremely low importance, medium importance and high importance according to the semantic importance scores and a preset dynamic threshold value; Mapping the two-dimensional probability mask to a high-resolution level image of the full-view pathology image through coordinates, dividing the high-resolution level image into grid tiles, and determining importance levels corresponding to the grid tiles; Performing a multi-branch adaptive differential compression strategy on grid tiles of different importance levels, wherein color parameters are extracted for the grid tiles of extremely low importance levels for miniature data block replacement; and generating and storing a differential pathology image file embedded with a differential compression index dictionary according to each grid block data after differential compression.
2. The adaptive differential compression method of full-field pathology images according to claim 1, wherein generating a two-dimensional probability mask comprising three levels of extremely low importance, medium importance and high importance according to the semantic importance score and a preset dynamic threshold value, comprises: Constructing a histogram of the semantic importance scores, and calculating an initial low score threshold and an initial high score threshold based on a multi-level threshold segmentation algorithm; Comparing the initial low-score threshold with a preset rigidity lower-limit constant, and selecting a smaller value as a final low-score threshold for cutting off and dividing the extremely low importance level; and calculating the proportion of the block data divided into high importance levels according to the initial high-score threshold, and if the proportion of the block data is lower than the preset high-frequency information retention limit proportion, forcibly reducing the initial high-score threshold to serve as a final high-score threshold based on the corresponding data percentile for dividing the high importance levels from the medium importance levels.
3. The full-view pathology image adaptive differential compression method according to claim 1, further comprising, before said performing a multi-branch adaptive differential compression strategy on grid tiles of different importance levels: generating a buffer transition zone at the space juncture of the grid blocks of the medium importance level and the high importance level; the method comprises the steps of detecting medium-importance grid tiles with high-importance grid tiles in the neighborhood of the medium-importance grid tiles, marking the medium-importance grid tiles as first-stage transition blocks, and performing multi-stage morphological expansion operation on the peripheral medium-importance areas by taking the first-stage transition blocks as starting points to generate multi-stage transition blocks to jointly form a buffer transition area with preset width.
4. A full-field pathology image adaptive differential compression method according to claim 3, wherein the multi-branch adaptive differential compression strategy is performed on grid tiles of different importance levels, further comprising: For grid tiles in the buffer transition zone, adopting a nonlinear compression quality factor based on a distance attenuation function to carry out coding, rewriting and compression; The nonlinear compression quality factor is dynamically calculated according to the topological network layer number distance of the high-importance grid block with the nearest current transition block, so that the compression quality factor is in nonlinear smooth attenuation from the lossless extremum of the high importance level to the lossy bottom of the medium importance level.
5. The adaptive differential compression method of full-field pathology images according to claim 1, characterized in that the micro-data block replacement is performed on the extracted color parameters of the grid tiles with extremely low importance level, and specifically comprises: Discarding the original pixel matrix data of the grid block with the extremely low importance level, and calculating and extracting arithmetic average values of a plurality of color channels of the original pixel matrix data; Generating a bypass identification head, splicing the bypass identification head and the arithmetic average value into a miniature data block with a fixed byte length, and writing the miniature data block into a physical storage stream to replace the traditional image discrete cosine transform coding data.
6. The adaptive differential compression method of full-field pathology images according to claim 1, wherein generating and storing differential pathology image files embedded with differential compression index dictionary according to each grid block data after differential compression specifically comprises: Repackaging the compressed data stream data into a pathological image format file containing file header metadata and differential data load; And embedding a differential compression index dictionary in the file header metadata, wherein the differential compression index dictionary records the coordinate positions of all grid tiles, the corresponding importance level strategies and the physical memory offset pointing to the differential data load.
7. The full-field pathology image adaptive differential compression method according to claim 6, further comprising a dynamic decoding rendering step of the pathology image at the client terminal: When a terminal requests to load a grid block of a viewport coordinate position, a scheduling analyzer queries a differential compression index dictionary in the differential pathological image file; If the target grid block is judged to be of an extremely low importance level, the analyzer triggers a hardware short circuit mechanism, a conventional image decompression pipeline of the central processing unit is disconnected, a color arithmetic average value is directly extracted from the miniature data block to serve as a global unified variable, a drawing instruction of a bottom layer graphic application program interface is generated, and the graphic processor directly renders the solid color geometric graphic primitive at a view port position of a corresponding coordinate.
8. An AI semantic importance-based full-field pathology image self-adaptive differential compression system, which is characterized by comprising: the pyramid image analysis module is used for acquiring a low-resolution level image of the full-view pathological image as a thumbnail; The semantic evaluation mask generation module is used for inputting the thumbnail into a pre-trained semantic evaluation model to output semantic importance scores of all areas and generating a two-dimensional probability mask containing three levels of extremely low importance, medium importance and high importance according to a dynamic threshold; The multi-branch differential coding module is used for mapping the two-dimensional probability mask coordinates into grid blocks of the high-resolution hierarchical image, extracting color parameters for extremely low-importance blocks to replace the color parameters, performing lossy compression for medium-importance blocks, performing lossless compression for high-importance blocks, generating smooth transition at medium-high junctions, and generating and storing differential pathology image files with differential compression index dictionaries.
9. An electronic device comprising a memory and a processor, wherein the memory has stored therein a computer program, the processor being arranged to run the computer program to perform the full-field pathology image adaptive differential compression method according to any one of claims 1 to 7.
10. A readable storage medium, characterized in that the readable storage medium has stored therein a computer program comprising program code for controlling a process to perform a process comprising the full-field pathology image adaptive differential compression method according to any one of claims 1 to 7.

Description

Full-field pathological image self-adaptive differential compression method based on AI semantic importance and application thereof Technical Field The application relates to the technical cross fields of digital pathology, image compression coding and artificial intelligence, in particular to a full-field pathology image self-adaptive differential compression method based on AI semantic importance and application thereof. Background Full field digital slicing (white SLIDE IMAGE, WSI) is a central data carrier for digital pathology, where the resolution of individual images is extremely large, typically on the order of 100000 x 100000 pixels, resulting in a single file volume as high as 1GB to 5GB. In clinical practice, hundreds of thousands of digital pathological slices are produced annually in a large and medium-sized hospital, which brings about extremely heavy construction and maintenance cost pressures to the storage servers (e.g., NAS/SAN architecture) of the hospital. Meanwhile, when application scenes such as remote pathological consultation are developed, the method is limited by network bandwidth bottleneck, and the ultra-large-volume WSI files are extremely easy to be blocked during transmission and loading, so that the clinical film reading efficiency is greatly affected. Existing WSI compression techniques typically employ a globally unified compression policy (e.g., unified compression based on algorithms such as JPEG, LZW, or JPEG 2000). The processing mode of the 'one-view kernel' has the contradiction that is difficult to overcome: on the one hand, if global lossless compression is adopted, the volume of the compressed file is still extremely large, and the storage and transmission pressure is difficult to effectively relieve; On the other hand, if global lossy compression based on high compression ratio is adopted, although the file volume can be obviously reduced, the loss of microscopic key diagnosis details such as tumor cell nucleolus, nuclear membrane and the like is inevitably caused, so that serious medical accident risks are brought. In fact, a conventional WSI pathology usually contains 60% to 80% of the blank glass background or normal interstitium and large adipose tissue without clinical diagnostic value, and the area truly containing core lesions (e.g. nuclear atypical, cell dense) is only a very small part. In the prior art, the semantic importance of the image cannot be recognized, and equal coding resources are allocated to all areas, so that great calculation load and storage space are wasted. Therefore, there is a need for an intelligent data compression and storage method, system and storage medium for ultra-large resolution pathology full-field digital slicing (WSI) to solve the problems of the prior art. Disclosure of Invention The embodiment of the application provides a full-view pathology image self-adaptive differential compression method based on AI semantic importance and application thereof, aiming at the problems that a global unified compression ratio is adopted in the prior art, so that a great amount of storage resources are wasted by a worthless background, core focus microcosmic details are extremely easy to lose due to high compression ratio, contradiction between extremely low storage occupation of an oversized pathology image and extremely high diagnosis fidelity requirement cannot be reconciled, and the like. The core technology of the invention is mainly a multi-path differential coding and decoding rendering system which is used for evaluating pixel-level semantic importance of a full-view pathological image based on a lightweight neural network, generating a grading mask, and adaptively implementing hardware bypass mean value replacement of a very low importance region, lossy compression of a common interstitial region and absolute lossless compression of a core high-risk region in different regions by combining a buffer smoothing algorithm. In a first aspect, the application provides a full-field pathological image self-adaptive differential compression method based on AI semantic importance, which comprises the following steps: Acquiring a low-resolution level image of the full-view pathology image as a thumbnail; Inputting the thumbnail into a pre-trained semantic evaluation model, and outputting semantic importance scores of all areas; Generating a two-dimensional probability mask comprising three levels of extremely low importance, medium importance and high importance according to the semantic importance score and a preset dynamic threshold; Mapping the two-dimensional probability mask to a high-resolution level image of the full-view pathology image through coordinates, dividing the high-resolution level image into grid blocks, and determining importance levels corresponding to the grid blocks; Performing a multi-branch adaptive differential compression strategy on grid tiles of different importance levels, wherein color parameters are extra