Medical lesion area automatic labeling method and system based on image processing

Abstract

The invention provides a medical lesion area automatic labeling method and system based on image processing, and relates to the technical field of medical image processing, and the method comprises the steps: collecting multi-source heterogeneous medical image data, and carrying out the standardized preprocessing of the multi-source heterogeneous medical image data; constructing a basic labeling model, performing primary basic labeling on the multi-source heterogeneous medical image data, and generating an initial medical lesion area labeling result; performing automatic correction and quality evaluation processing on the initial medical lesion area labeling result to generate a high-quality medical lesion area labeling result; the method comprises the following steps: constructing a fine labeling model, performing secondary fine labeling on a high-quality medical lesion area labeling result, and generating a final medical lesion area labeling result, so that full-process efficient automatic labeling of the medical lesion area can be realized, the labeling speed and precision are remarkably improved, and the medical lesion area labeling efficiency is improved. And seamless integration and traceable management of multi-source heterogeneous medical image data are realized, and finally reliable basic data are provided for medical image analysis.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, and in particular to a method and system for automatic annotation of medical lesion regions based on image processing. Background Technology

[0002] In the field of medical image processing, accurate identification and labeling of lesion areas is a key prerequisite for assisting clinical diagnosis, disease assessment, and treatment plan formulation.

[0003] However, traditional lesion annotation methods primarily rely on manual annotation of lesion areas. This method is not only time-consuming and labor-intensive but also prone to subjective bias and errors, especially when processing large amounts of heterogeneous medical image data from multiple sources. While advancements in artificial intelligence have led to the application of automated tools in image segmentation and object detection, these tools often lack a complete processing workflow, failing to effectively integrate data upload, preprocessing, and automatic annotation. This results in low annotation efficiency, unstable accuracy, and difficulty in establishing a closed-loop mechanism for continuous optimization. Furthermore, existing methods often neglect the importance of data augmentation and anomaly detection when dealing with the complexities of medical images, such as noise interference, inconsistent formats, and anomalous samples. This further limits the reliability and traceability of the annotation results.

[0004] Therefore, it is necessary to provide an automatic annotation method and system for medical lesion regions based on image processing to solve the above-mentioned technical problems. Summary of the Invention

[0005] To address the aforementioned technical problems, this invention provides an automatic annotation method and system for medical lesion regions based on image processing, which solves the technical problems of low efficiency, unstable accuracy, and lack of closed-loop optimization in existing medical image annotation methods.

[0006] The present invention provides an automatic annotation method for medical lesion regions based on image processing, the method comprising: Collect multi-source heterogeneous medical image data and perform standardized preprocessing on the multi-source heterogeneous medical image data; A basic annotation model is constructed to perform basic annotation on the multi-source heterogeneous medical image data, generating initial medical lesion region annotation results. Automated correction and quality assessment are performed on the initial medical lesion region annotation results to generate high-quality medical lesion region annotation results; A fine-grained annotation model is constructed, and the high-quality medical lesion region annotation results are further finely annotated to generate the final medical lesion region annotation results.

[0007] Preferably, the standardization preprocessing of the multi-source heterogeneous medical image data specifically includes: Image processing tools were used to perform grayscale conversion, pixel value normalization, and Gaussian denoising on the multi-source heterogeneous medical image data; The multi-source heterogeneous medical image data is subjected to data augmentation processing by calling a data augmentation library. The data augmentation processing includes at least random cropping, angle rotation, horizontal flipping, and adaptive brightness adjustment. The isolated forest algorithm is used to identify and filter abnormal samples in the multi-source heterogeneous medical image data.

[0008] Preferably, the construction of the basic annotation model, which involves performing a basic annotation on the multi-source heterogeneous medical image data to generate initial medical lesion region annotation results, specifically includes: The basic annotation model includes an image segmentation sub-model and an object detection sub-model; If the multi-source heterogeneous medical image data is random medical image data, the image segmentation sub-model uses SAM to perform zero-shot segmentation on the random medical image data to generate a zero-shot segmentation mask; if the multi-source heterogeneous medical image data is specified medical image data, the image segmentation sub-model uses a combination of nnU-Net, 3D-UNet, and MONAI frameworks to perform high-precision segmentation on the organ and lesion boundaries of the specified medical image data to generate a high-precision segmentation mask; the zero-shot segmentation mask or the high-precision segmentation mask is used as the image segmentation mask; The target detection sub-model uses the YOLOv8 and Detectron2 frameworks to perform lesion target detection on the multi-source heterogeneous medical image data after image segmentation, generating lesion target bounding boxes. The image segmentation confidence threshold and the target detection confidence threshold are set to perform quality screening on the image segmentation mask and the lesion target bounding box, respectively. The image segmentation mask with a confidence level not lower than the image segmentation confidence threshold and the lesion target bounding box with a confidence level not lower than the target detection confidence threshold are retained. The image segmentation mask, the bounding box of the lesion target, and the associated medical lesion region information are encapsulated in a structured format to generate the initial medical lesion region annotation result.

[0009] Preferably, the step of performing automated correction and quality assessment on the initial medical lesion region annotation results to generate high-quality medical lesion region annotation results specifically includes: For the initial medical lesion region annotation results, the image segmentation mask with breaks is detected by the connected component analysis algorithm, and the anatomical prior knowledge is called to filter the lesion target bounding box that exceeds the reasonable anatomical range. The image segmentation mask with breaks and the lesion target bounding box that exceeds the reasonable anatomical range are marked as low-quality medical lesion region annotation data. Automated correction is performed on the low-quality medical lesion region annotation data. For image segmentation masks with breaks, interpolation optimization is performed using a continuous annotation algorithm for adjacent image slices. For lesion target bounding boxes that exceed the reasonable anatomical range, the Canny edge detection algorithm is called to correct the lesion edges. A multi-dimensional quality assessment system is constructed to calculate the medical lesion quality index of the corrected initial medical lesion region annotation result. The medical lesion quality index includes the mask Dice similarity coefficient and mask intersection-union ratio calculated based on the image segmentation mask, and the average accuracy of the bounding box calculated based on the lesion target bounding box. If the mask Dice similarity coefficient, the mask intersection-union ratio, and the bounding box average precision are all higher than the coefficient index threshold, the intersection-union ratio index threshold, and the precision index threshold, respectively, then the initial medical lesion region annotation result is determined to be the high-quality medical lesion region annotation result.

[0010] Preferably, the step of constructing a fine-grained annotation model, which performs secondary fine-grained annotation on the high-quality medical lesion region annotation results to generate the final medical lesion region annotation results, specifically includes: High-quality medical lesion region annotation features are extracted from the high-quality medical lesion region annotation results. The high-quality medical lesion region annotation features include image modality lesion features and annotation-related lesion features. A multimodal feature fusion network is constructed, comprising a cross-modal attention module and a dynamic weight fusion module. The cross-modal attention module calculates the interaction weights between the image modal lesion features and the labeled associated lesion features, and performs interactive fusion of the image modal lesion features and the labeled associated lesion features to output cross-modal medical lesion region annotation features. The dynamic weight fusion module combines the medical image modal characteristics in the historical medical image annotation dataset, calculates the lesion annotation contribution coefficients corresponding to the image modal lesion features and the labeled associated lesion features, and outputs weighted fused medical lesion annotation features. The cross-modal medical lesion region annotation features and the weighted fused medical lesion annotation features are summarized to generate the final medical lesion region annotation result.

[0011] Preferably, the cross-modal attention module calculates the interaction weights between the image modal lesion features and the labeled associated lesion features, and performs interactive fusion of the image modal lesion features and the labeled associated lesion features to output cross-modal medical lesion region labeled features, specifically including: The cross-modal attention module uses a scaled dot product method to calculate the interaction weight matrix between the image modal lesion features and the labeled associated lesion features. The corresponding calculation formula is as follows: In the formula, Indicates the modal equilibrium coefficient; The feature matrix representing the characteristics of lesions in imaging modalities; The feature matrix represents the labeled features associated with the lesion. The transpose of the matrix; Indicates the feature dimension scaling factor; Represents the normalization function; Feature matrix representing the characteristics of lesions in imaging modalities Feature matrix associated with labeled lesion features Dot product of the transpose of a matrix; This indicates that a convolution operation is performed on the feature matrix of the lesion features of the imaging modality; The image modal lesion features and the labeled associated lesion features are interactively fused to output a feature matrix of cross-modal medical lesion region labeled features. The corresponding calculation formula is as follows: In the formula, This indicates that the matrix is ​​multiplied element by element.

[0012] Preferably, the dynamic weight fusion module combines the modal characteristics of medical images in the historical medical image annotation dataset to calculate the contribution coefficient of the lesion annotations corresponding to the lesion features of the image modality and the lesion features associated with the annotations, specifically including: A modal lesion matching preprocessing operation is performed on the historical medical image annotation dataset to construct a targeted historical sample library; Based on the improved Spearman rank correlation coefficient algorithm, the correlation strength between the image modal lesion features and the labeled associated lesion features and the lesion labeling results in the target historical sample library is calculated. The corresponding calculation formula is as follows: In the formula, n represents the number of samples in the target historical sample library; Represents the rank sequence value of the image modal lesion features in the i-th sample; The average value of the grade sequence values ​​representing the characteristics of the lesion in the imaging modality; This represents the rank sequence value of the lesion labeling result in the i-th sample; This represents the average value of the rank sequence values ​​of the lesion annotation results in the target historical sample library; The correlation strength between the labeled lesion features and the lesion labeling results in the targeted historical sample library. The calculation method is the same as above; Will and After normalization, the lesion annotation contribution coefficients of the lesion features of the image modality were obtained respectively. The contribution coefficient of lesion annotation to lesion features associated with the annotation. .

[0013] An automatic medical lesion region annotation system based on image processing, the system comprising: The data processing module is used to collect multi-source heterogeneous medical image data and perform standardized preprocessing on the multi-source heterogeneous medical image data; The basic annotation module is used to construct a basic annotation model, perform basic annotation on the multi-source heterogeneous medical image data, and generate initial medical lesion area annotation results. The correction and evaluation module is used to perform automated correction and quality evaluation on the initial medical lesion region annotation results to generate high-quality medical lesion region annotation results. The fine annotation module is used to construct a fine annotation model, perform secondary fine annotation on the high-quality medical lesion area annotation results, and generate the final medical lesion area annotation results.

[0014] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor performs the steps of the image processing-based automatic annotation method for medical lesion regions as described in any of the preceding claims.

[0015] A readable storage medium storing a computer program, which, when executed by a processor, is used to implement the steps of the image processing-based automatic annotation method for medical lesion regions as described above.

[0016] Compared with related technologies, the automatic annotation method and system for medical lesion regions based on image processing provided by this invention has the following beneficial effects: This invention acquires multi-source heterogeneous medical image data and performs standardized preprocessing on it; constructs a basic annotation model to perform initial basic annotation on the multi-source heterogeneous medical image data, generating initial medical lesion region annotation results; performs automated correction and quality assessment on the initial medical lesion region annotation results to generate high-quality medical lesion region annotation results; constructs a fine annotation model to perform secondary fine annotation on the high-quality medical lesion region annotation results, generating final medical lesion region annotation results. This enables efficient and automated annotation of medical lesion regions throughout the entire process, significantly improving annotation speed and accuracy, and achieving seamless integration and traceable management of multi-source heterogeneous medical image data, ultimately providing reliable basic data for medical image analysis.

[0017] This invention effectively eliminates differences and noise interference from medical images of different modalities and formats by performing standardized preprocessing on acquired multi-source heterogeneous medical image data, including grayscale conversion, pixel value normalization, Gaussian denoising, data augmentation, and anomaly filtering using the Isolation Forest algorithm. This solves the problem of the difficulty in uniformly processing traditional multi-source data. The invention constructs a basic annotation model containing image segmentation and target detection sub-models. For random medical image data, it uses zero-sample segmentation; for specified medical image data, it uses multi-frame combination high-precision segmentation. Then, lesion target bounding boxes are generated through lesion target detection, and a confidence threshold is used for filtering to adapt to different types of image annotation needs, improving the efficiency and adaptability of initial annotation. This invention uses a connected component analysis algorithm to detect broken image segmentation masks, calls anatomical prior knowledge to filter lesion target bounding boxes that exceed reasonable anatomical ranges, accurately identifies low-quality medical lesion region annotation data, and uses interpolation optimization and Canny edge detection to automatically correct low-quality medical lesion region annotation data. Simultaneously, a multi-dimensional quality assessment system is constructed to strictly screen high-quality medical lesion region annotation results, significantly reducing the cost of manual intervention and significantly improving the accuracy and consistency of annotation results. This invention constructs a multimodal feature fusion network containing a cross-modal attention module and a dynamic weight fusion module to extract two types of lesion features from high-quality medical lesion region annotation results. It achieves feature interaction fusion and weighted fusion, and optimizes the feature contribution coefficient by combining historical medical image annotation datasets, which significantly improves the accuracy of secondary fine annotation and ensures that the final medical lesion region annotation results are more in line with the actual clinical application needs. Attached Figure Description

[0018] Figure 1 A flowchart illustrating an automatic annotation method for medical lesion regions based on image processing, provided in an embodiment of the present invention; Figure 2 A system block diagram of an automatic medical lesion region annotation system based on image processing provided in an embodiment of the present invention; Figure 3 This is a schematic diagram of the hardware structure of an electronic device provided in an embodiment of the present invention. Detailed Implementation

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0020] like Figure 1The diagram shown is a flowchart of an automatic annotation method for medical lesion regions based on image processing, provided in an embodiment of the present invention. Figure 1 The execution entity of the method shown can be a software and / or hardware device. The execution entity of this application can include, but is not limited to, at least one of the following: user equipment, network equipment, etc. User equipment can include, but is not limited to, computers, smartphones, personal digital assistants (PDAs), and the aforementioned electronic devices. Network equipment can include, but is not limited to, a single network server, a server group consisting of multiple network servers, or a cloud based on cloud computing consisting of a large number of computers or network servers. Cloud computing is a type of distributed computing, consisting of a super virtual computer composed of a group of loosely coupled computers. This embodiment does not limit this. Steps S1 to S4 are detailed as follows: S1, Collect multi-source heterogeneous medical image data and perform standardized preprocessing on the multi-source heterogeneous medical image data; The standardization preprocessing of the multi-source heterogeneous medical image data specifically includes: Image processing tools were used to perform grayscale conversion, pixel value normalization, and Gaussian denoising on the multi-source heterogeneous medical image data; The multi-source heterogeneous medical image data is subjected to data augmentation processing by calling a data augmentation library. The data augmentation processing includes at least random cropping, angle rotation, horizontal flipping, and adaptive brightness adjustment. The isolated forest algorithm is used to identify and filter abnormal samples in the multi-source heterogeneous medical image data.

[0021] Grayscale conversion converts multi-channel or color information from heterogeneous medical image data from multiple sources into single-channel grayscale information. By unifying the visual representation dimension of the images, it eliminates processing adaptation problems caused by differences in the number of channels between different modalities. Pixel value normalization maps the pixel values ​​of all images to a preset unified numerical range based on the parameter differences of image acquisition equipment, such as the different grayscale ranges of ultrasound equipment. This avoids overemphasis on high-level pixel features due to differences in pixel value magnitude. Gaussian denoising constructs a filter kernel conforming to a Gaussian distribution to smoothly suppress random noise in the images caused by equipment noise, electromagnetic interference, etc., reducing noise interference while preserving key details of the lesion area and improving the image signal-to-noise ratio.

[0022] Data augmentation addresses the limitations of limited medical image sample quantity and monotonous scene representation. It extracts effective local regions containing lesions or target organs from complete images through random cropping, enriching the diversity of local features. Angle rotation and horizontal flipping simulate image angle changes caused by patient positioning variations in clinical settings, enhancing robustness to images in different positions. Adaptive brightness adjustment addresses the issue of uneven image brightness caused by different acquisition devices and scanning parameters. By dynamically adjusting pixel brightness distribution, it ensures image brightness remains within a suitable and uniform range for recognition, preventing brightness differences from masking lesion features.

[0023] The anomaly identification and filtering operation targets anomalous samples in multi-source heterogeneous medical image data that have failed to be acquired, have corrupted formats, or have incorrect annotations. It utilizes the characteristic of isolated forests, which quantifies the isolation degree of samples by constructing multiple isolated trees, to quickly filter out and remove anomalous samples that differ significantly from normal image features. This avoids anomalous samples interfering with the model annotation accuracy and ensures the reliability of the overall annotation process.

[0024] In practical applications, multi-source heterogeneous medical image data comes from different hospitals and different models of acquisition equipment, resulting in significant differences in format, pixel range, noise type, and sample quality. Grayscale conversion eliminates the differences in the number of channels between different modalities of images, unifying the image processing dimensions. Pixel value normalization avoids interference from differences in the magnitude of pixel values ​​between different devices. Gaussian denoising suppresses random noise such as device noise and electromagnetic interference, preserving lesion details while improving image clarity. Data augmentation addresses the limited clinical sample size and differences in body position by simulating different acquisition scenarios to enrich sample diversity. The Isolation Forest algorithm quickly identifies and filters abnormal samples that failed to acquire or have corrupted formats, ensuring the validity of data entering the annotation process.

[0025] S2, Construct a basic annotation model, perform basic annotation on the multi-source heterogeneous medical image data, and generate initial medical lesion area annotation results; The construction of the basic annotation model involves performing a basic annotation on the multi-source heterogeneous medical image data to generate initial medical lesion region annotation results, specifically including: The basic annotation model includes an image segmentation sub-model and an object detection sub-model; If the multi-source heterogeneous medical image data is random medical image data, the image segmentation sub-model uses SAM to perform zero-shot segmentation on the random medical image data to generate a zero-shot segmentation mask; if the multi-source heterogeneous medical image data is specified medical image data, the image segmentation sub-model uses a combination of nnU-Net, 3D-UNet, and MONAI frameworks to perform high-precision segmentation on the organ and lesion boundaries of the specified medical image data to generate a high-precision segmentation mask; the zero-shot segmentation mask or the high-precision segmentation mask is used as the image segmentation mask; The target detection sub-model uses the YOLOv8 and Detectron2 frameworks to perform lesion target detection on the multi-source heterogeneous medical image data after image segmentation, generating lesion target bounding boxes. The image segmentation confidence threshold and the target detection confidence threshold are set to perform quality screening on the image segmentation mask and the lesion target bounding box, respectively. The image segmentation mask with a confidence level not lower than the image segmentation confidence threshold and the lesion target bounding box with a confidence level not lower than the target detection confidence threshold are retained. The image segmentation mask, the bounding box of the lesion target, and the associated medical lesion region information are encapsulated in a structured format to generate the initial medical lesion region annotation result.

[0026] SAM is a general image segmentation model that achieves general segmentation capabilities for any image through a three-module architecture of "image encoder - cue encoder - mask decoder," and is used for zero-shot segmentation of random medical image data. nnU-Net is a medical image segmentation framework, 3D-UNet is a 3D medical image segmentation network, and the MONAI framework is an open-source deep learning framework for medical images. The combination of these three is used for high-precision segmentation of specified medical image data. YOLOv8 is a real-time object detection framework used for lesion detection after image segmentation. It can quickly locate the lesion position within the segmented region and generate initial lesion bounding boxes. Detectron2 is an open-source object detection and instance segmentation framework used for accuracy correction of lesion bounding boxes. Associated medical lesion region information refers to auxiliary metadata directly associated with the lesion region corresponding to the image segmentation mask and lesion bounding box, used to supplement the clinical background, image source, and basic attributes of the lesion.

[0027] Understandably, the basic annotation model, through the collaborative efforts of the image segmentation sub-model and the target detection sub-model, respectively undertakes the functions of defining the lesion area and accurately locating the lesion target, forming an annotation logic of "first define the region, then define the location".

[0028] Random medical image data refers to general medical images without pre-defined image modalities or examination sites, such as routine outpatient images randomly received in clinical practice. Due to the high modality diversity and variable application scenarios of this type of data, SAM is used to circumvent the limitations of training separately for specific modalities. Leveraging its segmentation adaptation capabilities for general medical images, it quickly generates zero-shot segmentation masks to meet basic segmentation needs in multiple scenarios. Specified medical image data refers to images with pre-defined modalities, such as lung CT scans and brain MRI scans, targeting specific examination sites. This type of data requires higher segmentation accuracy; therefore, a combination of nnU-Net, 3D-UNet, and MONAI frameworks is used. Through the complementary advantages of multiple frameworks, organ contours and lesion boundary details are accurately captured, generating high-precision segmentation masks to meet the precise segmentation needs of specific clinical scenarios.

[0029] Image segmentation masks are used to distinguish different regions based on differences in pixel values. They typically use preset high-brightness pixel values ​​to mark the determined lesion area and target organ range, and low-brightness pixel values ​​to mark the background area, intuitively presenting the spatial location and morphological outline of the lesion in the image.

[0030] The combination of YOLOv8 and Detectron2 frameworks further locates lesions within the area defined by the image segmentation mask. YOLOv8 is responsible for quickly outputting the initial lesion target bounding box, while Detectron2 optimizes the accuracy of the lesion target bounding box, especially correcting the detection results of small lesions and lesions with blurred edges, to ensure the accuracy of lesion target location labeling.

[0031] Image segmentation confidence threshold and target detection confidence threshold are reliability assessment indicators preset based on clinical annotation standards and training experience of the basic annotation model. They are used to screen high-confidence annotation results. Segmentation masks and lesion target bounding boxes with confidence scores below the corresponding confidence thresholds are removed, and only results that meet the confidence threshold requirements are retained, ensuring the basic reliability of the initial medical lesion region annotation results.

[0032] Structured format encapsulation is used to integrate and store image segmentation masks, lesion target bounding boxes, and related medical lesion region information according to a preset format, so that the initial annotation results contain both visual annotation information and are associated with complete image metadata, providing data format support for the collaboration of the entire annotation process.

[0033] S3, perform automated correction and quality assessment on the initial medical lesion region annotation results to generate high-quality medical lesion region annotation results; The automatic correction and quality assessment of the initial medical lesion region annotation results to generate high-quality medical lesion region annotation results specifically includes: For the initial medical lesion region annotation results, the image segmentation mask with breaks is detected by the connected component analysis algorithm, and the anatomical prior knowledge is called to filter the lesion target bounding box that exceeds the reasonable anatomical range. The image segmentation mask with breaks and the lesion target bounding box that exceeds the reasonable anatomical range are marked as low-quality medical lesion region annotation data. Automated correction is performed on the low-quality medical lesion region annotation data. For image segmentation masks with breaks, interpolation optimization is performed using a continuous annotation algorithm for adjacent image slices. For lesion target bounding boxes that exceed the reasonable anatomical range, the Canny edge detection algorithm is called to correct the lesion edges. A multi-dimensional quality assessment system is constructed to calculate the medical lesion quality index of the corrected initial medical lesion region annotation result. The medical lesion quality index includes the mask Dice similarity coefficient and mask intersection-union ratio calculated based on the image segmentation mask, and the average accuracy of the bounding box calculated based on the lesion target bounding box. If the mask Dice similarity coefficient, the mask intersection-union ratio, and the bounding box average precision are all higher than the coefficient index threshold, the intersection-union ratio index threshold, and the precision index threshold, respectively, then the initial medical lesion region annotation result is determined to be the high-quality medical lesion region annotation result.

[0034] The connected component analysis algorithm identifies and marks sets of connected pixels with the same pixel value by traversing the pixel region of the image segmentation mask. If the area to be covered by the same lesion is segmented into multiple discontinuous connected pixel blocks, i.e., there are interval regions with pixel values ​​as background, the image segmentation mask is determined to be broken, thereby accurately locating the regional integrity defects caused by model segmentation deviations in the initial annotation.

[0035] Anatomical prior knowledge is a knowledge base built upon clinical anatomical standards and the laws of human physiological structure. It covers the normal anatomical location, morphological range, and adjacent relationships of different organs and tissues. For example, lung lesions can only exist in the thoracic lung lobe region and cannot appear in the abdominal liver region. By comparing the coordinates of the lesion target bounding box with the reasonable anatomical range of the corresponding organ, lesion target bounding boxes that are outside the range are filtered out. This eliminates obvious mislabeling from a clinical logic perspective, ensuring that the labeling results conform to common sense about human physiological structure.

[0036] The continuous annotation algorithm for adjacent image slices is designed for the characteristic that medical images are mostly in the form of sequential slices. Its core logic is that lesions usually have spatial continuity in adjacent image slices. For example, the morphology and location of the same tumor should gradually transition in consecutive CT slices, rather than abruptly or breaking off. Therefore, by extracting the complete lesion contour features in adjacent slices of the break mask, interpolation calculations, such as linear interpolation and bilinear interpolation, are used to complete the pixel information at the break point. This allows the corrected segmentation mask to restore the spatial continuity of the lesion region, solving the problem of incomplete regions caused by the break in a single slice segmentation.

[0037] The Canny edge detection algorithm employs multiple processing steps, such as Gaussian smoothing for noise reduction, image gradient calculation, non-maximum suppression, and double-threshold edge connectivity, to accurately locate the true edges of lesion regions. For lesion target bounding boxes that exceed the reasonable anatomical range, the algorithm identifies the actual edge pixel coordinates of the lesion and then adjusts the vertex coordinates of the bounding box to ensure that the corrected lesion target bounding box fits the true edge of the lesion, eliminating annotation errors caused by the positional offset of the lesion target bounding box.

[0038] The Dice similarity coefficient of the mask is an indicator that measures the degree of overlap between the image segmentation mask and the real lesion region. It is suitable for evaluating the segmentation quality of small lesions or irregularly shaped lesions. The mask intersection-over-union ratio (IoU) is a segmentation accuracy evaluation indicator used to measure the overall matching degree between the segmented region and the real region, comprehensively evaluating the image segmentation mask quality from the perspectives of region integrity and background purity. The average bounding box precision is an indicator that evaluates the accuracy of lesion target bounding box localization, effectively reflecting the reliability of the target detection sub-model in annotating the lesion location.

[0039] The coefficient threshold, intersection-union ratio threshold, and accuracy threshold are reliability judgment standards preset based on a large number of high-quality clinical annotation samples. Their values ​​are set with reference to the minimum requirements of clinical annotation accuracy. For example, the mask Dice similarity coefficient threshold usually meets the error tolerance of clinical diagnosis in determining the extent of lesions. Only when all indicators of the corrected annotation result are higher than the corresponding threshold can it be judged to meet the high-quality standard, ensuring that the output high-quality medical lesion area annotation results have the accuracy and reliability required for clinical application.

[0040] By employing the above methods, low-quality medical lesion region annotation data in the initial medical lesion region annotation results can be accurately located, and targeted corrections can be made using automated means. This effectively eliminates biases and defects in the initial annotation, significantly reduces the cost of manual intervention, and improves the reliability and efficiency of medical lesion region annotation.

[0041] S4. Construct a fine-grained annotation model, perform secondary fine-grained annotation on the high-quality medical lesion region annotation results, and generate the final medical lesion region annotation results.

[0042] The construction of the fine-grained annotation model involves performing secondary fine-grained annotation on the high-quality medical lesion region annotation results to generate the final medical lesion region annotation results. Specifically, this includes: High-quality medical lesion region annotation features are extracted from the high-quality medical lesion region annotation results. The high-quality medical lesion region annotation features include image modality lesion features and annotation-related lesion features. A multimodal feature fusion network is constructed, comprising a cross-modal attention module and a dynamic weight fusion module. The cross-modal attention module calculates the interaction weights between the image modal lesion features and the labeled associated lesion features, and performs interactive fusion of the image modal lesion features and the labeled associated lesion features to output cross-modal medical lesion region annotation features. The dynamic weight fusion module combines the medical image modal characteristics in the historical medical image annotation dataset, calculates the lesion annotation contribution coefficients corresponding to the image modal lesion features and the labeled associated lesion features, and outputs weighted fused medical lesion annotation features. The cross-modal medical lesion region annotation features and the weighted fused medical lesion annotation features are summarized to generate the final medical lesion region annotation result.

[0043] Among them, high-quality medical lesion region annotation features are information carriers extracted from high-quality medical lesion region annotations that have been corrected and evaluated. They include two types of key sub-features: one is image modality lesion features, which are features directly related to the modal attributes of the medical image itself, such as the tissue density features of CT images, reflecting the differences in the physical morphology and structure of lesions under different modalities; the other is annotation-related lesion features, which are auxiliary features directly related to the annotation information, such as the association attributes of the preliminarily determined lesion type, supplementing the clinical context information of the annotation results.

[0044] Multimodal feature fusion network solves the problem of insufficient single feature representation ability through the synergistic effect of cross-modal attention module and dynamic weight fusion module, breaks down the information barrier between image modal lesion features and labeled associated lesion features, and achieves complementary advantages.

[0045] The cross-modal attention module automatically identifies information that plays a key role in lesion annotation by calculating the interaction weights between lesion features in the image modality and lesion features associated with the annotation. For example, the strong correlation between lesion edge features and annotation location information in a certain modality. Then, the key information is integrated through interactive fusion to output cross-modal medical lesion region annotation features. This feature can simultaneously retain the detailed information of the image modality and the contextual information associated with the annotation, avoiding the omission of key annotation basis by a single feature.

[0046] The core advantage of the dynamic weighted fusion module lies in its adaptability to the characteristics of different medical imaging modalities. Based on historical medical image annotation datasets encompassing multimodal images and corresponding fine-grained annotation results, it statistically analyzes the impact of image modality lesion features and annotation-related lesion features on annotation accuracy under different modalities, obtaining a lesion annotation contribution coefficient. This coefficient quantifies the importance of image modality lesion features and annotation-related lesion features. For example, in brain MRI images, image modality lesion features contribute more to lesion boundary annotation; in abdominal ultrasound images, annotation-related lesion features contribute more to the auxiliary determination of lesion type. Then, based on the lesion annotation contribution coefficient, the two types of features are weighted and fused to output weighted fused medical lesion annotation features, avoiding the problem of insufficient fusion adaptability caused by fixed weights.

[0047] Historical medical image annotation datasets serve as the foundation for dynamic weight calculation. They are preprocessed using modal lesion matching to construct a targeted historical sample library, ensuring that the statistically derived lesion annotation contribution coefficients conform to actual clinical annotation patterns and improving the rationality of weight allocation.

[0048] Cross-modal medical lesion region annotation features and weighted fusion medical lesion annotation features are the two core outputs of the multimodal feature fusion network. The former focuses on mining the interaction and correlation information between the two types of features, while the latter focuses on weight adaptation based on modal characteristics. The sum of the two can form a fusion feature with stronger representation capabilities. Finally, the final medical lesion region annotation result is generated based on this fusion feature, ensuring that the secondary fine annotation has both the accuracy of multi-feature collaboration and the flexibility to adapt to different modalities.

[0049] In practical applications, high-quality medical lesion region annotation features are extracted specifically based on the characteristics of multimodal images. For CT images, the focus is on extracting lesion features related to tissue density differences; for MRI, the focus is on signal intensity changes; and for ultrasound, the focus is on echogenicity. Annotated lesion features are simultaneously associated with information such as the anatomical location of the lesion and its relationship with adjacent organs. A cross-modal attention module enhances the interaction of key features, such as the interaction weight between prominent lesion signal features and the proximity of ventricles in brain MRI annotations. The dynamic weight fusion module relies on historical medical image annotation datasets covering samples from multiple hospitals and devices, optimizing the allocation of lesion annotation contribution coefficients under different scenarios. The final aggregated medical lesion region annotation results are more aligned with clinical diagnostic needs, reducing the cost of manual correction.

[0050] The cross-modal attention module calculates the interaction weights between the image modal lesion features and the labeled associated lesion features, and performs interactive fusion of the image modal lesion features and the labeled associated lesion features to output cross-modal medical lesion region labeled features, specifically including: The cross-modal attention module uses a scaled dot product method to calculate the interaction weight matrix between the image modal lesion features and the labeled associated lesion features. The corresponding calculation formula is as follows: In the formula, Indicates the modal equilibrium coefficient; The feature matrix representing the characteristics of lesions in imaging modalities; The feature matrix represents the labeled features associated with the lesion. The transpose of the matrix; Indicates the feature dimension scaling factor; Represents the normalization function; Feature matrix representing the characteristics of lesions in imaging modalities Feature matrix associated with labeled lesion features Dot product of the transpose of a matrix; This indicates that a convolution operation is performed on the feature matrix of the lesion features of the imaging modality; The image modal lesion features and the labeled associated lesion features are interactively fused to output a feature matrix of cross-modal medical lesion region labeled features. The corresponding calculation formula is as follows: In the formula, This indicates that the matrix is ​​multiplied element by element.

[0051] The modal balance coefficient is used to balance the contribution ratio of image modal lesion features and labeled associated lesion features in the interaction weight calculation. This avoids the situation where one type of feature is too strong, such as prominent MRI image signal features, and obscures the clinical value of another type of feature, such as the relative position labeling information of lesions and surrounding organs. This ensures that both types of features can play an effective role in the interaction weight matrix, which is in line with the need for medical lesion labeling to take into account both image details and clinical context.

[0052] Understandably, the cross-modal attention module uses a scaled dot product approach to adapt to the high dimensionality and high information density of medical image features, accurately capturing the correlation strength between image modal lesion features and labeled associated lesion features, thus avoiding the omission of key information in medical images by traditional calculation methods.

[0053] The dynamic weight fusion module combines the modal characteristics of medical images in the historical medical image annotation dataset to calculate the contribution coefficient of the lesion annotations corresponding to the lesion features of the image modality and the lesion features associated with the annotations. Specifically, this includes: A modal lesion matching preprocessing operation is performed on the historical medical image annotation dataset to construct a targeted historical sample library; Based on the improved Spearman rank correlation coefficient algorithm, the correlation strength between the image modal lesion features and the labeled associated lesion features and the lesion labeling results in the target historical sample library is calculated. The corresponding calculation formula is as follows: In the formula, n represents the number of samples in the target historical sample library; Represents the rank sequence value of the image modal lesion features in the i-th sample; The average value of the grade sequence values ​​representing the characteristics of the lesion in the imaging modality; This represents the rank sequence value of the lesion labeling result in the i-th sample; This represents the average value of the rank sequence values ​​of the lesion annotation results in the target historical sample library; The correlation strength between the labeled lesion features and the lesion labeling results in the targeted historical sample library. The calculation method is the same as above; Will and After normalization, the lesion annotation contribution coefficients of the lesion features of the image modality were obtained respectively. The contribution coefficient of lesion annotation to lesion features associated with the annotation. .

[0054] Among them, the preprocessing operation of modality lesion matching is used to realize the correspondence between image modality and lesion type in historical medical image annotation dataset. That is, according to the goal of the current medical lesion region annotation task, samples with the same modality and the same lesion type are selected from historical data, and samples with modality confusion and lesion type irrelevant are removed to avoid interference from irrelevant data.

[0055] The targeted historical sample library is a specialized sample collection formed after modal lesion matching processing. Its significant feature is its scenario-specificity. Compared with unfiltered general historical medical image annotation datasets, the targeted historical sample library only contains samples that are highly relevant to the current annotation task, and centrally reflects the correlation between image modal lesion features, annotation-related lesion features, and lesion annotation results under specific modalities.

[0056] The improved Spearman rank correlation coefficient algorithm is optimized for the specific characteristics of medical image features, building upon the traditional Spearman rank correlation coefficient. It adapts to the ranked representation of modal lesion features in medical images and the ranked description of lesion annotation results. By calculating the correlation between the rank sequence of modal lesion features and lesion annotation results, it measures the consistency between changes in modal lesion features and changes in lesion annotation results.

[0057] The rank sequence value is a feature and annotation result expression form adapted to the improved Spearman rank correlation coefficient algorithm. It transforms continuous image modal lesion features and discrete annotated associated lesion features into ordered rank values, solving the problem that it is difficult to directly calculate the correlation between medical image features and annotation results due to differences in data types, and ensuring the feasibility and accuracy of correlation strength calculation.

[0058] The lesion annotation contribution coefficient is the final product after correlation strength calculation and normalization. Among them, the lesion annotation contribution coefficient of image modality lesion features is used to quantify the supporting weight of image features themselves to annotation accuracy, and the lesion annotation contribution coefficient of annotation-related lesion features is used to quantify the supporting weight of annotation context information to annotation accuracy. The two types of coefficients together provide a clear basis for the weighting operation of the dynamic weight fusion module, ensuring that the dynamic weight fusion module can reasonably allocate the proportion of image modality lesion features and annotation-related lesion features in the fusion process according to historical data patterns, and ensure that the output weighted fused medical lesion annotation features are adapted to the annotation characteristics of different modalities of images.

[0059] like Figure 2 The diagram shown is a system block diagram of an automatic medical lesion region annotation system based on image processing provided in an embodiment of the present invention. The system includes: The data processing module is used to collect multi-source heterogeneous medical image data and perform standardized preprocessing on the multi-source heterogeneous medical image data; The basic annotation module is used to construct a basic annotation model, perform basic annotation on the multi-source heterogeneous medical image data, and generate initial medical lesion area annotation results. The correction and evaluation module is used to perform automated correction and quality evaluation on the initial medical lesion region annotation results to generate high-quality medical lesion region annotation results. The fine annotation module is used to construct a fine annotation model, perform secondary fine annotation on the high-quality medical lesion area annotation results, and generate the final medical lesion area annotation results.

[0060] Figure 2 The apparatus of the illustrated embodiment can be used to perform corresponding actions. Figure 1 The steps in the method embodiments shown are implemented in a similar manner and have similar technical effects, and will not be repeated here.

[0061] An electronic device includes a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, the processor performs the steps of the image processing-based automatic annotation method for medical lesion regions as described in any of the preceding claims.

[0062] like Figure 3The diagram shown is a hardware structure schematic of an electronic device according to an embodiment of the present invention. The electronic device 30 includes: a processor 31, a memory 32, and a computer program; wherein... The memory 32 is used to store the computer program, and the memory may also be flash memory. The computer program is, for example, an application program or functional module that implements the above method.

[0063] Processor 31 is configured to execute the computer program stored in the memory to implement the various steps performed by the device in the above method. For details, please refer to the relevant descriptions in the preceding method embodiments.

[0064] Alternatively, the memory 32 can be either standalone or integrated with the processor 31.

[0065] When the memory 32 is a device independent of the processor 31, the device may further include: Bus 33 is used to connect the memory 32 and the processor 31.

[0066] A readable storage medium storing a computer program, which, when executed by a processor, is used to implement the steps of the image processing-based automatic annotation method for medical lesion regions as described above.

[0067] The readable storage medium can be a computer storage medium or a communication medium. A communication medium includes any medium that facilitates the transfer of computer programs from one location to another. A computer storage medium can be any available medium accessible to a general-purpose or special-purpose computer. For example, a readable storage medium is coupled to a processor, enabling the processor to read information from and write information to the readable storage medium. Of course, the readable storage medium can also be a component of the processor. The processor and the readable storage medium can reside in an Application-Specific Integrated Circuit (ASIC). Alternatively, the ASIC can be located in a user equipment. Of course, the processor and the readable storage medium can also exist as discrete components in a communication device. The readable storage medium can be a read-only memory (ROM), random access memory (RAM), CD-ROM, magnetic tape, floppy disk, and optical data storage device, etc.

[0068] The present invention also provides a program product including executable instructions stored in a readable storage medium. At least one processor of the device can read the executable instructions from the readable storage medium, and the at least one processor executes the executable instructions to cause the device to implement the methods provided in the various embodiments described above.

[0069] In the embodiments of the above-described device, it should be understood that the processor can be a Central Processing Unit (CPU), or other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), etc. The general-purpose processor can be a microprocessor or any conventional processor. The steps of the method disclosed in this invention can be directly manifested as execution by a hardware processor, or execution by a combination of hardware and software modules within the processor.

[0070] Through the above embodiments, this invention achieves efficient and automatic annotation of medical lesion regions throughout the entire process by collecting multi-source heterogeneous medical image data and performing standardized preprocessing on the data; constructing a basic annotation model to perform initial basic annotation on the data, generating initial medical lesion region annotation results; performing automated correction and quality assessment on the initial medical lesion region annotation results to generate high-quality annotation results; and constructing a fine annotation model to perform secondary fine annotation on the high-quality results, generating final annotation results. This significantly improves annotation speed and accuracy, enables seamless integration and traceable management of multi-source heterogeneous medical image data, and ultimately provides reliable basic data for medical image analysis.

[0071] This invention effectively eliminates differences and noise interference from medical images of different modalities and formats by performing standardized preprocessing on acquired multi-source heterogeneous medical image data, including grayscale conversion, pixel value normalization, Gaussian denoising, data augmentation, and anomaly filtering using the Isolation Forest algorithm. This solves the problem of the difficulty in uniformly processing traditional multi-source data. The invention constructs a basic annotation model containing image segmentation and target detection sub-models. For random medical image data, it uses zero-sample segmentation; for specified medical image data, it uses multi-frame combination high-precision segmentation. Then, lesion target bounding boxes are generated through lesion target detection, and a confidence threshold is used for filtering to adapt to different types of image annotation needs, improving the efficiency and adaptability of initial annotation. This invention uses a connected component analysis algorithm to detect broken image segmentation masks, calls anatomical prior knowledge to filter lesion target bounding boxes that exceed reasonable anatomical ranges, accurately identifies low-quality medical lesion region annotation data, and uses interpolation optimization and Canny edge detection to automatically correct low-quality medical lesion region annotation data. Simultaneously, a multi-dimensional quality assessment system is constructed to strictly screen high-quality medical lesion region annotation results, significantly reducing the cost of manual intervention and significantly improving the accuracy and consistency of annotation results. This invention constructs a multimodal feature fusion network containing a cross-modal attention module and a dynamic weight fusion module to extract two types of lesion features from high-quality medical lesion region annotation results. It achieves feature interaction fusion and weighted fusion, and optimizes the feature contribution coefficient by combining historical medical image annotation datasets, which significantly improves the accuracy of secondary fine annotation and ensures that the final medical lesion region annotation results are more in line with the actual clinical application needs.

[0072] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. An automatic labeling method of a medical lesion region based on image processing, characterized by, The method comprises: Collecting multi-source heterogeneous medical image data and performing standardized preprocessing on the multi-source heterogeneous medical image data; Building a basic annotation model, performing a basic annotation on the multi-source heterogeneous medical image data, and generating an initial medical lesion region annotation result; Performing automatic correction and quality evaluation processing on the initial medical lesion region annotation result to generate a high-quality medical lesion region annotation result; Building a fine annotation model, performing a secondary fine annotation on the high-quality medical lesion region annotation result, and generating a final medical lesion region annotation result.

2. The method for automatic labeling of medical lesion regions based on image processing according to claim 1, characterized in that, The standardized preprocessing of the multi-source heterogeneous medical image data specifically comprises: Performing grayscale, pixel value normalization and Gaussian denoising operations on the multi-source heterogeneous medical image data using an image processing tool; Performing data enhancement processing on the multi-source heterogeneous medical image data by calling a data enhancement library, wherein the data enhancement processing at least includes random cropping, angle rotation, horizontal flipping and brightness adaptive adjustment; Performing anomaly sample identification and filtering operations on the multi-source heterogeneous medical image data using an isolation forest algorithm. 3.The method of claim 1, wherein, The basic annotation model comprises an image segmentation sub-model and a target detection sub-model; If the multi-source heterogeneous medical image data is random medical image data, the image segmentation sub-model performs zero-shot segmentation on the random medical image data using SAM to generate a zero-shot segmentation mask; if the multi-source heterogeneous medical image data is specified medical image data, the image segmentation sub-model performs high-precision segmentation on the organ and lesion boundary of the specified medical image data using a combination of nnU-Net, 3D-UNet and MONAI framework to generate a high-precision segmentation mask; the zero-shot segmentation mask or the high-precision segmentation mask is used as an image segmentation mask; The target detection sub-model uses YOLOv8 and Detectron2 framework to perform lesion target detection on the multi-source heterogeneous medical image data after image segmentation to generate a lesion target bounding box; Set image segmentation confidence threshold and target detection confidence threshold to quality screen the image segmentation mask and the lesion target bounding box respectively, retain the image segmentation mask whose confidence is not lower than the image segmentation confidence threshold, and the lesion target bounding box whose confidence is not lower than the target detection confidence threshold; The image segmentation mask, the lesion target bounding box and the associated medical lesion region information are packaged in a structured format to generate the initial medical lesion region annotation result. The automatic correction and quality evaluation processing of the initial medical lesion region annotation result specifically comprises:

4. The method of claim 1, wherein the method further comprises: ​ For the initial medical lesion region annotation result, a connected component analysis algorithm is used to detect the image segmentation mask that is broken, and anatomical prior knowledge is called to filter the lesion target bounding box that exceeds the reasonable anatomical range, and the image segmentation mask that is broken and the lesion target bounding box that exceeds the reasonable anatomical range are marked as low-quality medical lesion region annotation data; For the low-quality medical lesion region annotation data, automatic correction is performed, for the image segmentation mask that is broken, a continuous annotation algorithm of adjacent image slices is used for interpolation optimization, and for the lesion target bounding box that exceeds the reasonable anatomical range, a Canny edge detection algorithm is called for lesion edge correction; A multi-dimensional quality evaluation system is constructed, and medical lesion quality indicators of the corrected initial medical lesion region annotation result are calculated, the medical lesion quality indicators including a mask Dice similarity coefficient and a mask intersection over union calculated based on the image segmentation mask, and a bounding box average precision calculated based on the lesion target bounding box; If the mask Dice similarity coefficient, the mask intersection over union and the bounding box average precision are higher than a coefficient indicator threshold, an intersection over union indicator threshold and a precision indicator threshold respectively, it is determined that the initial medical lesion region annotation result is the high-quality medical lesion region annotation result.

5. The method of claim 1, wherein the method further comprises: The fine annotation model is constructed, the high-quality medical lesion region annotation result is further fine annotated to generate a final medical lesion region annotation result, specifically including: High-quality medical lesion region annotation features are extracted from the high-quality medical lesion region annotation result, the high-quality medical lesion region annotation features including image modality lesion features and annotation associated lesion features; A multi-modal feature fusion network including a cross-modal attention module and a dynamic weight fusion module is constructed, the cross-modal attention module calculating interaction weights of the image modality lesion features and the annotation associated lesion features, and interactively fusing the image modality lesion features and the annotation associated lesion features to output cross-modal medical lesion region annotation features, and the dynamic weight fusion module combining medical image modality characteristics in a historical medical image annotation dataset to statistically calculate lesion annotation contribution coefficients corresponding to the image modality lesion features and the annotation associated lesion features, and output weight fusion medical lesion annotation features after weighting, and the cross-modal medical lesion region annotation features and the weight fusion medical lesion annotation features are summarized to generate the final medical lesion region annotation result.

6. The method of claim 5, wherein the method further comprises: The cross-modal attention module calculates interaction weights of the image modality lesion features and the annotation associated lesion features, and interactively fuses the image modality lesion features and the annotation associated lesion features to output cross-modal medical lesion region annotation features, specifically including: The cross-modal attention module adopts a scaled dot product method to calculate an interaction weight matrix of the image modality lesion feature and the labeled associated lesion feature The corresponding calculation formula is as follows: wherein, denotes a modal balance coefficient; denotes a feature matrix of image modality lesion characteristics; denotes a feature matrix of labeled associated lesion characteristics denotes a transpose matrix of the feature matrix of labeled associated lesion characteristics denotes a feature dimension scaling factor; denotes a normalization function; denotes a feature matrix of image modality lesion characteristics denotes a point multiplication operation of the feature matrix of image modality lesion characteristics and a transpose matrix of the feature matrix of labeled associated lesion characteristics denotes a point multiplication operation of the feature matrix of image modality lesion characteristics and a transpose matrix of the feature matrix of labeled associated lesion characteristics denotes a convolution operation on the feature matrix of image modality lesion characteristics; And the image modal lesion characteristics and the labeled associated lesion characteristics are interactively fused to output a feature matrix of the cross-modal medical lesion region labeling characteristics The corresponding calculation formula is as follows: wherein denotes the element-wise multiplication of matrices.

7. The method of claim 5, wherein the method further comprises: The dynamic weight fusion module combines medical image modality characteristics in a historical medical image annotation dataset to statistically calculate lesion annotation contribution coefficients corresponding to the image modality lesion features and the annotation associated lesion features, specifically including: A preprocessing operation of modality lesion matching is performed on the historical medical image annotation dataset to construct a targeted historical sample library; calculate the correlation strength of the image modality lesion feature and the lesion annotation result in the targeted historical sample library based on an improved Spearman rank correlation coefficient algorithm, and calculate the correlation strength of the labeled associated lesion feature and the lesion annotation result in the targeted historical sample library The corresponding calculation formula is as follows: wherein n represents the number of samples in the targeted historical sample library; represents the rank sequence value of the image modality lesion feature in the i-th sample; represents the average value of the rank sequence value of the image modality lesion feature; represents the rank sequence value of the lesion annotation result in the i-th sample; represents the average value of the rank sequence value of the lesion annotation result in the targeted historical sample library; the correlation strength of the lesion annotation features with the lesion annotation results in the target historical sample library The calculation method is the same as above; will be described in detail below with reference to the accompanying drawings. with After normalization, the lesion annotation contribution coefficients of the lesion features of the image modalities are obtained respectively with the lesion annotation contribution coefficients of the lesion features associated with the annotation .

8. The automatic labeling system of medical lesion region based on image processing, applied to the automatic labeling method of medical lesion region based on image processing as claimed in any one of claims 1-7, characterized in that, The system includes: The data processing module is configured to collect multi-source heterogeneous medical image data and perform standardization preprocessing on the multi-source heterogeneous medical image data. The basic annotation module is configured to construct a basic annotation model, perform one-time basic annotation on the multi-source heterogeneous medical image data, and generate an initial medical lesion region annotation result. The correction evaluation module is configured to perform automatic correction and quality evaluation processing on the initial medical lesion region annotation result, and generate a high-quality medical lesion region annotation result. The fine annotation module is configured to construct a fine annotation model, perform two-time fine annotation on the high-quality medical lesion region annotation result, and generate a final medical lesion region annotation result.

9. An electronic device comprising a memory and a processor, said memory having stored therein a computer program, characterized in that, When the processor runs the computer program stored in the memory, the processor executes the steps of the image processing-based automatic medical lesion region annotation method according to any one of claims 1-7.

10. A readable storage medium, in which a computer program is stored, characterized in that, The computer program is executed by the processor to implement the steps of the image processing-based automatic medical lesion region annotation method according to any one of claims 1-7.

Images