A processing method and system for spinal image labeling

Abstract

This invention relates to the field of image processing and discloses a processing method and system for spinal image labeling. The system achieves automated spinal image labeling through a spinal image acquisition module, an image preprocessing module, an AI-based vertebral recognition and localization module, and a label generation and output module. The core of this invention lies in the introduction of an AI-driven quality verification and control module. This module intelligently scores and grades the preprocessed images, initiates a dual-path differentiated AI processing flow for images with critical quality, and performs intelligent decision-making by quantitatively evaluating the differences between the two paths and fusing the original quality scores. It triggers an early warning when there are high-risk contradictions in the identification results. This invention improves the robustness of intelligent spinal image analysis software in complex real-world scenarios, the reliability of output results, and the safety of clinical use, representing high-quality medical AI application software development.

Description

Technical Field

[0001] This invention relates to the field of image processing, and more specifically to a processing method and system for spinal image identification. Background Technology

[0002] Accurate identification and localization of spinal images play a crucial role in medical analysis and evaluation. Traditional spinal image analysis primarily relies on manual interpretation by physicians, involving visual observation of X-ray, CT, or MRI images and manual marking of the location and type of each vertebra. This method is not only inefficient but also highly dependent on the physician's personal experience, potentially leading to inconsistencies between different physicians or even within the same physician on multiple interpretations, resulting in subjective differences and insufficient reproducibility.

[0003] In recent years, with the development of computer vision and deep learning technologies, several automatic spinal image recognition methods based on models such as Convolutional Neural Networks (CNNs) have emerged. These methods can, to some extent, detect and locate vertebrae, helping to reduce the workload of manual labor. However, existing AI-based spinal recognition methods still have significant shortcomings: First, as an AI application, they lack robustness and adaptive processing capabilities to fluctuations in the quality of input images (such as noise and blur). Second, and more importantly, such systems lack intelligent verification and quality control mechanisms for the reliability of their own AI recognition results. When the model makes errors in the face of complex or low-quality images, it cannot detect and warn of these errors, posing a risk of outputting incorrect results without realizing it. This severely restricts the practical application value of such AI-optimized operating systems in scenarios with high reliability requirements, such as clinical diagnosis. Summary of the Invention

[0004] The purpose of this invention is to provide a processing method and system for spinal image identification, thereby solving at least one of the above-mentioned technical problems.

[0005] The objective of this invention can be achieved through the following technical solutions:

[0006] A processing system for spinal image labeling, comprising:

[0007] A spinal image acquisition module is used to acquire raw medical images containing the spine; the raw medical images include at least one of X-ray films, CT scan images, or MRI images.

[0008] The image preprocessing module is used to preprocess the original medical images to obtain standardized spinal images; the preprocessing includes: image size normalization, contrast enhancement, noise filtering and image rotation correction;

[0009] The quality verification and control module is used to assess the quality of the standardized spinal images that have passed quality verification, and to take corresponding processing strategies based on the assessment results. It also performs consistency verification on the identification results and outputs a quality warning when the identification risk exceeds the threshold.

[0010] The vertebral identification and localization module is used to detect vertebrae in the standardized spinal images and output the vertebral location information and type labels.

[0011] The identifier generation module is used to generate vertebral identifiers based on the location information and type label. The identifiers include vertebral serial number, bounding box, segment color differentiation, and medical parameter annotations.

[0012] The labeling output module is used to overlay the vertebral labels onto the standardized spinal image and output a spinal image with labels.

[0013] As a further technical solution, the vertebral recognition and localization module is implemented using a deep learning model, which includes a convolutional neural network (CNN), a recurrent neural network (RNN), or a Transformer. The location information output by the deep learning model includes the coordinates of the vertebral bounding box and the coordinates of the center point, and the type labels include cervical vertebrae C1-C7, thoracic vertebrae T1-T12, lumbar vertebrae L1-L5, and sacral vertebrae.

[0014] As a further technical solution, the quality verification and control module specifically includes:

[0015] The preliminary screening subunit is used to calculate the comprehensive quality score of each standardized spinal image based on multiple preset quality indicators, and output three types of results based on the comparison of the comprehensive quality score with preset thresholds: images that have passed the preliminary screening, images that have failed the preliminary screening, and images in a borderline state.

[0016] The dual-processing subunit inputs the standardized spinal images that have passed the initial screening to the vertebral identification and localization module to obtain the first identification result and the first confidence level.

[0017] The standardized spinal images of the critical state are input into the vertebral recognition and localization module using two differentiated strategies to obtain a second recognition result, a second confidence level, a third recognition result, and a third confidence level.

[0018] For images that clearly fail the initial screening, a manual review instruction is triggered;

[0019] The consistency assessment subunit is used to calculate the difference between the second identification result and the third identification result. The difference is calculated based on the difference in vertebral center coordinates, the deviation in intervertebral space height, the difference in spinal curvature, and the regional weighting coefficient.

[0020] The early warning unit is used to output quality early warning information when the fusion decision value F calculated based on the comprehensive quality score and the difference value exceeds a preset limit. The early warning information includes the risk of vertebral sequence disorder or local identification deviation.

[0021] As a further technical solution, the difference value of the consistency evaluation subunit is calculated using the following formula:

[0022] Through the formula: The difference value was calculated. ;

[0023] Among them, the first one is obtained respectively. Difference in center coordinates of vertebrae , No. The height deviation between the intervertebral segment and the adjacent vertebrae Overall curvature of the spine and the first Regional weighting coefficient of vertebrae ;

[0024] This is a sequence consistency correction function; when there are duplicates or misordered vertebral type label sequences, The value should be between 0.5 and 1.0, otherwise... ; The spinal curvature in the second identification result. The third identification result is the spinal curvature.

[0025] As a further technical solution, the differentiation strategy includes fine-tuning of preprocessing parameters, switching of model inference modes, or integration of multiple models.

[0026] As a further technical solution, the preset quality indicators include image clarity, noise intensity, and vertebral region integrity;

[0027] The process of calculating the comprehensive quality score in the preliminary screening subunit is as follows:

[0028] The comprehensive quality score S, normalized to the 0-1 range, is calculated by weighted summation based on image clarity, noise intensity, and vertebral region integrity.

[0029] The process by which the preliminary screening subunit outputs three types of results based on the comprehensive quality score S is as follows:

[0030] When the overall quality score is higher than or equal to the first threshold, it is determined that the preliminary screening has been clearly passed;

[0031] When the overall quality score is lower than or equal to the second threshold, it is determined that the initial screening has been clearly failed.

[0032] When the overall quality score is between the first threshold and the second threshold, it is determined to be a critical state image;

[0033] Wherein, the second threshold is less than the first threshold.

[0034] As a further technical solution, the formula for calculating the fusion decision value F is: ;

[0035] in, The difference value calculated for the consistency evaluation subunit. The scaling factor is the normalization coefficient. The quality sensitivity coefficient, and , To prevent the division of zero decimals, S represents the overall quality score;

[0036] For critical state images, when the fusion decision value F > 1, quality warning information is output.

[0037] As a further technical solution, the preprocessing parameter fine-tuning in the differentiation strategy is based on dynamic adjustment of the local quality features of the critical state image, specifically including:

[0038] The dual-processing subunit also includes a local quality feature analysis subunit, which is used to identify low-quality regions and high-interest vertebral segments in standardized spinal images and output a region quality score.

[0039] The first processing path for critical state images is configured as follows:

[0040] For low-quality regions corresponding to high-interest vertebral segments, super-resolution reconstruction based on generative adversarial networks is used to improve local detail resolution.

[0041] For low-quality areas that are not of interest, only bilateral filtering is used for basic noise reduction;

[0042] The second processing path for the same critical state image is configured as follows:

[0043] For low-quality regions corresponding to high-interest vertebral segments, a multi-scale fusion algorithm is used to enhance edge information;

[0044] For low-quality regions that are not of interest, noise interference is suppressed by adaptive threshold pruning;

[0045] Furthermore, the two processing paths employ vertebral recognition models with different initial weights.

[0046] A processing method for spinal image identification includes the following steps:

[0047] S1. Acquire raw spinal images and generate standardized spinal images through the image preprocessing module;

[0048] S2. Use the vertebral identification and localization module to detect the location and type of vertebrae in standardized spinal images;

[0049] S3. Based on the detection results, generate vertebral labels through the label generation module and overlay them onto the standardized spinal image output;

[0050] S4. Perform quality verification on the images before output: Through preliminary screening, distinguish between images that have passed, are in a critical state, and have failed. Perform dual-path differential processing and consistency assessment on images in a critical state. If the difference exceeds the limit, an early warning will be triggered.

[0051] The beneficial effects of this invention are:

[0052] (1) In addition to the core recognition process, the present invention constructs a quality verification and control module. It realizes the internal loop verification through a three-layer progressive mechanism of preliminary screening, dual-path processing and consistency assessment. Specifically, firstly, a comprehensive quality score is calculated based on multiple indicators such as image clarity, noise intensity and vertebral region integrity to achieve preliminary classification of the input image. For images with critical quality, a dual-path differentiation strategy is automatically started. Two sets of recognition results are generated by using different preprocessing parameters, model inference modes and even heterogeneous model architectures in parallel. On this basis, the consistency difference value of the multi-dimensional features of vertebral center coordinate difference, intervertebral space height deviation and spinal curvature difference is fused and combined with the comprehensive quality score to construct a fusion decision value. Thus, based on objective quantitative indicators, suspicious results are automatically identified and warnings are triggered, thereby reducing the risk of misidentification caused by image quality fluctuations.

[0053] (2) This invention introduces a local quality feature analysis mechanism in the dual-processing subunit. This mechanism can accurately identify low-quality regions and high-interest vertebral segments in standardized spinal images. Based on this analysis, two independent processing paths are dynamically configured for critical state images: The first path uses super-resolution reconstruction technology based on generative adversarial network (GAN) to significantly improve local detail resolution for low-quality regions of high-interest vertebral segments with a 4x magnification factor; while for similar problems in non-interest regions, only parameter-optimized bilateral filtering is used for basic noise reduction to balance processing efficiency; The second path uses a multi-scale fusion algorithm to enhance edge information for the same high-interest regions; and for non-interest regions, noise is suppressed by adaptive threshold pruning; Through a differentiated strategy based on semantic attention, it is ensured that limited computing resources are prioritized for key regions that have the greatest impact on recognition accuracy, thus achieving the optimal balance between overall efficiency and local accuracy. Attached Figure Description

[0054] The invention will now be further described with reference to the accompanying drawings.

[0055] Figure 1 This is a system logic block diagram of the present invention. Detailed Implementation

[0056] 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.

[0057] Please see Figure 1 As shown, the present invention is a processing system for spinal image identification, comprising:

[0058] A spinal image acquisition module is used to acquire raw medical images containing the spine; the raw medical images include at least one of X-ray films, CT scan images, or MRI images.

[0059] The image preprocessing module is used to preprocess the original medical images to obtain standardized spinal images; the preprocessing includes: image size normalization, contrast enhancement, noise filtering and image rotation correction;

[0060] The quality verification and control module is used to perform quality assessment on the standardized spinal images, and take corresponding processing strategies based on the assessment results. It performs consistency verification on the identification results and outputs a quality warning when the identification risk exceeds the threshold.

[0061] The vertebral identification and localization module is used to detect vertebrae in the standardized spinal images that have passed quality verification, and output the vertebral location information and type label;

[0062] The identifier generation module is used to generate vertebral identifiers based on the location information and type label. The identifiers include vertebral serial number, bounding box, segment color differentiation, and medical parameter annotations. The medical parameters include at least one of intervertebral disc height and spinal curvature.

[0063] The labeling output module is used to overlay the vertebral labels onto the standardized spinal image and output a spinal image with labels.

[0064] The vertebral identification and localization module is implemented using a deep learning model, which includes a convolutional neural network (CNN), a recurrent neural network (RNN), or a Transformer. The location information output by the deep learning model includes the coordinates of the vertebral bounding box and the coordinates of the center point. The type labels include cervical vertebrae C1-C7, thoracic vertebrae T1-T12, lumbar vertebrae L1-L5, and sacral vertebrae.

[0065] In this embodiment, a complete fully automated spinal image labeling system was constructed, fundamentally changing the traditional model that relied on manual annotation, achieving an order-of-magnitude improvement in processing efficiency. Through a standardized preprocessing workflow, it effectively overcomes image inconsistencies caused by differences in equipment, imaging parameters, or patient positioning, laying a reliable foundation for subsequent accurate identification. More importantly, by employing an advanced deep learning model, it can not only locate vertebrae but also automatically identify and label their specific anatomical types, such as C1-C7, T1-T12, etc., providing directly usable standardized data for subsequent structured analysis and report generation. This avoids labeling errors or sequence disorder that are difficult to completely eliminate in manual identification, ensuring the objectivity and consistency of the output results. Specifically…

[0066] First, the spinal image acquisition module is responsible for accessing raw medical images from multiple modalities. Then, the image preprocessing module performs a series of standardization operations, including image size normalization, contrast enhancement, noise filtering, and rotation correction, to eliminate image differences caused by non-anatomical structural factors and output standardized spinal images. Next, the vertebral recognition and localization module, as the intelligent core, uses deep learning models such as Convolutional Neural Networks (CNN), Recurrent Neural Networks (RNN), or Transformers to analyze the standardized images, simultaneously outputting precise location information and type labels for each vertebra, such as bounding box and center point coordinates. The label generation module then generates visual labels containing serial numbers, bounding boxes, segmental color distinctions, and medical parameters based on the recognition results. Finally, the label output module seamlessly overlays the generated labels onto the original images, directly outputting the final usable labeled spinal image, completing the processing.

[0067] The quality verification and control module specifically includes:

[0068] The preliminary screening subunit is used to calculate the comprehensive quality score of each standardized spinal image based on multiple preset quality indicators, and output three types of results based on the comparison of the comprehensive quality score with preset thresholds: images that have passed the preliminary screening, images that have failed the preliminary screening, and images in a borderline state.

[0069] The dual-processing subunit inputs the standardized spinal images that have passed the initial screening to the vertebral identification and localization module to obtain the first identification result and the first confidence level.

[0070] The standardized spinal images of the critical state are input into the vertebral recognition and localization module using two differentiated strategies to obtain a second recognition result, a second confidence level, a third recognition result, and a third confidence level.

[0071] For images that clearly fail the initial screening, a manual review instruction is triggered;

[0072] The consistency assessment subunit is used to calculate the difference between the second identification result and the third identification result. The difference is calculated based on the difference in vertebral center coordinates, the deviation in intervertebral space height, the difference in spinal curvature, and the regional weighting coefficient.

[0073] The early warning unit is used to output quality early warning information when the fusion decision value F calculated based on the comprehensive quality score and the difference value exceeds a preset limit. The early warning information includes the risk of vertebral sequence disorder or local identification deviation.

[0074] The difference value of the consistency assessment subunit is calculated using the following formula:

[0075] Through the formula: The difference value was calculated. ;

[0076] Among them, the first one is obtained respectively. Difference in center coordinates of vertebrae , No. The height deviation between the intervertebral segment and the adjacent vertebrae Overall curvature of the spine and the first Regional weighting coefficient of vertebrae ;

[0077] It should be noted that the regional weighting coefficient It is based on the type of vertebrae, such as cervical vertebrae. ,thoracic lumbar spine Because the cervical spine has a complex structure and is difficult to identify, it is given a higher weight; the overall curvature of the spine is: the average curvature of the midline curve obtained by fitting the center point of each vertebra or the Cobb angle formed by the lower endplate of cervical C2 and the upper endplate of thoracic T12.

[0078] This is a sequence consistency correction function; when there are duplicates or misordered vertebral type label sequences, The value should be between 0.5 and 1.0, otherwise... ; The spinal curvature is represented by the dimensionless second identification result. The third identification result is a dimensionless measurement of spinal curvature.

[0079] The specific explanation of the above formula is as follows:

[0080] The design goal of this formula is to transform and amplify the subtle differences between the two recognition results into a single quantitative indicator that can intuitively reflect the reliability of the recognition result. Its significance lies in integrating inconsistencies across three dimensions: local, sequential, and global, rather than a simple averaging. Specifically, This represents a weighted assessment of the differences between local location and structure. It measures the difference in the judgment of the position of the same vertebra, and is the most basic test of the consistency of positioning; It measures the difference in the judgment of the relationship between vertebrae. Even if the location of a single vertebra is slightly off, if its relative positional relationship with the adjacent vertebrae is consistent, then the deviation is acceptable and can be keenly detected by sequential identification errors. It combines the absolute locational differences and relative structural differences of a single vertebra to form a comprehensive local score for the consistency of vertebral identification.

[0081] This indicates an amplification of the severity of logical errors in the sequence. Specifically designed to amplify low-level but serious errors, such as duplicate, skipped, or reversed vertebral numbers. Take a value between 0.5 and 1.0; this term acts as a multiplicative amplifier, and in the event of a sequence error, it will amplify the maximum value of the core term by 1.5 to 2 times, clearly indicating that sequence disorder is more serious than slight positional shifts.

[0082] This represents the bottom-line guarantee of global morphological consistency. The overall curvature of the spine is one of the gold standards for assessing spinal health. This is done independently of the max operation and is added together. Even if the local positioning difference of each vertebra is very small, as long as there is a significant difference in the overall spinal morphology obtained by the two methods, the difference value will still be large, which can ensure the correctness of the global morphology.

[0083] By combining weighted summation, multiplicative amplification, and additive superposition strategies, the differences in four levels—vertebral location, intervertebral relationship, sequence logic, and overall morphology—are integrated into a single indicator, providing a reliable input for subsequent decision-making.

[0084] The differentiation strategy includes:

[0085] For example, fine-tuning the preprocessing parameters for strong noise reduction or weak enhancement versus weak noise reduction or strong enhancement;

[0086] For example, enabling or disabling the model inference mode switching for the Dropout layer;

[0087] Or, for example, multi-model integration such as parallel processing of CNN and Transformer.

[0088] In this embodiment, to address the reliability issues in automated systems caused by poor input image quality, an embedded, multi-level quality verification and control framework is introduced. Its core effect is to endow the system with the ability to self-doubt and self-verify. Through initial screening, it can automatically identify borderline images with questionable quality and initiate a more cautious dual-path processing flow for these images. The consistency assessment mechanism quantifies the differences between the two recognition results, transforming the uncertainty in the recognition process into an objective, quantifiable difference value. Ultimately, it no longer simply outputs a potentially erroneous result but can judge the reliability of its own output and trigger warnings or manual review when the confidence level is insufficient, greatly improving safety in complex real-world clinical environments and effectively preventing misidentification. Specifically…

[0089] The preliminary screening subunit in the quality verification and control module first assesses the quality of each standardized image and categorizes them into three types based on preset rules: clearly passed, clearly failed, and borderline. The dual-processing subunit then employs differentiated strategies for different image categories: high-quality images are directly identified; borderline images are processed in parallel using two differentiated strategies—one of preprocessing parameter fine-tuning, model inference mode switching, or multi-model ensemble—to generate two sets of identification results; low-quality images are directly submitted for manual intervention. The consistency assessment subunit specifically processes the dual-path results of borderline images, using a combined approach that considers the vertebral center coordinate differences. Intervertebral space height deviation Regional weights and differences in spinal curvature formula To calculate the difference value This formula can keenly detect key errors such as vertebral sequence disorder and positioning deviation; finally, when the fusion decision value calculated from the comprehensive quality score and the difference value exceeds the preset limit, the early warning unit outputs a quality warning, completing the entire verification closed loop.

[0090] The preset quality indicators include image clarity, noise intensity, and vertebral region integrity.

[0091] The process of calculating the comprehensive quality score in the preliminary screening subunit is as follows:

[0092] The comprehensive quality score S, normalized to the 0-1 range, is calculated by weighted summation based on image clarity, noise intensity, and vertebral region integrity.

[0093] The process by which the preliminary screening subunit outputs three types of results based on the comprehensive quality score S is as follows:

[0094] When the overall quality score is higher than or equal to the first threshold, it is determined that the preliminary screening has been clearly passed;

[0095] When the overall quality score is lower than or equal to the second threshold, it is determined that the initial screening has been clearly failed.

[0096] When the overall quality score is between the first threshold and the second threshold, it is determined to be a critical state image;

[0097] Wherein, the second threshold is less than the first threshold.

[0098] The formula for calculating the fusion decision value F is: ;

[0099] in, The difference value calculated for the consistency evaluation subunit. The scaling factor is the normalization coefficient. The quality sensitivity coefficient, and , To prevent the division of zero decimals, S represents the overall quality score;

[0100] For critical state images, when the fusion decision value F > 1, a quality warning is output. It should be noted that the scale normalization coefficient... By calculating the difference values ​​on the training set The mean value is obtained; the quality sensitivity coefficient is... The value range is [1.5, 2.5]; the value of the decimal ε for preventing the division of zero is 1 × 10⁻⁶. -6 ;

[0101] In this embodiment, firstly, the calculation method for the comprehensive quality score S is determined. Specifically, it involves weighting and summing three key indicators—image sharpness, noise intensity, and vertebral region integrity—and normalizing them, transforming a complex image quality assessment problem into a scalar value between 0 and 1, thus achieving precise quantification of quality. Based on this, a formula for calculating the fusion decision value F is then proposed. In this model, the numerator represents the difference value indicating inconsistency in identification. The denominator is the overall score representing the image quality level. The function after coefficient adjustment; this formula cleverly constructs a ratio relationship: when image quality... At lower values, the denominator is smaller, thus making the same difference value appear smaller. A larger fusion decision value F is calculated, making it easier to trigger an early warning; conversely, when S is higher, the denominator is larger, and the system has a higher tolerance for differences. By comparing this fusion value F with a fixed threshold of 1, the above adaptive decision-making logic is achieved. On the one hand, it changes the simple approach of traditional quality control systems that only set a fixed threshold, and deeply fuses information from two dimensions: the quality level of the image itself and the inherent consistency of the recognition result, thereby making a more reasonable final judgment;

[0102] Specifically, it can dynamically adjust its tolerance for recognition differences: for images of poor quality, even if there are large inconsistencies in the recognition results, it may be normal and will be given higher tolerance; while for high-quality images, the recognition results must be highly consistent. This adaptive mechanism greatly reduces false alarms and false negatives when dealing with boundary cases. For example, it provides unnecessary warnings for poor-quality images and ignores errors that should not occur in high-quality images, making quality warnings more accurate and reliable.

[0103] The fine-tuning of preprocessing parameters in the differentiation strategy is based on dynamic adjustment of local quality features of critical state images, specifically including:

[0104] The dual-path processing subunit also includes a local quality feature analysis subunit, which is used to identify low-quality regions such as local blurring and high-noise areas in standardized spinal images, as well as high-interest vertebral segments such as C1-C7 and L1-L5, and output a regional quality score.

[0105] The first processing path for critical state images is configured as follows:

[0106] For low-quality regions corresponding to high-interest vertebral segments, super-resolution reconstruction based on generative adversarial networks (GANs), such as the ESRGAN model, is used with a 4x magnification factor to improve local detail resolution.

[0107] For low-quality areas that are not of interest, only bilateral filtering is used for basic noise reduction; for example, spatial sigma = 0.1, grayscale sigma = 0.2.

[0108] The second processing path for the same critical state image is configured as follows:

[0109] For low-quality regions corresponding to high-interest vertebral segments, a multi-scale fusion algorithm is used to enhance edge information; for example, the multi-scale fusion algorithm fuses Laplacian pyramid features at three scales.

[0110] For low-quality regions that are not of interest, noise interference is suppressed by adaptive thresholding; for example, adaptive thresholding is based on the mean gray level of the region ± 2 times the standard deviation.

[0111] Furthermore, the two processing paths employ vertebral recognition models with different initialization weights; for example, the first path model is initialized based on a normal image training set, while the second path model is initialized based on a noisy image training set.

[0112] In this embodiment, a specific method for fine-tuning the preprocessing parameters is given. First, a local quality feature analysis subunit is introduced. This unit not only identifies low-quality regions in the image, such as blurred or high-noise areas, but also locates high-interest vertebral segments, such as C1-C7 and L1-L5, based on anatomical knowledge, and generates regional quality scores. Based on this analysis, two independent processing paths are dynamically configured for the same critical state image.

[0113] The first approach is to use high-fidelity detail restoration for low-quality regions of high-interest vertebral segments by employing super-resolution reconstruction based on generative adversarial networks (GANs); and to use only lightweight bilateral filtering for basic noise reduction for low-quality regions of non-interest areas.

[0114] The second approach employs a multi-scale fusion algorithm to enhance the edge and texture information of high-interest regions, while using adaptive threshold cropping to suppress noise in non-interest regions. Furthermore, both approaches utilize recognition models initialized on different datasets, such as normal image sets and noisy image sets, to further amplify the processing diversity at the algorithmic level. This ensures that the two sets of results submitted to the consistency evaluation subunit have sufficient diversity to facilitate the detection of potential inconsistencies.

[0115] Through the above process, differentiated processing strategies can be advanced from macro-level strategy selection to micro-level pixel-level optimization, achieving precise allocation of computing resources and a significant improvement in the accuracy of key anatomical structure recognition. Traditional image enhancement methods typically apply the same algorithm to the entire image, which may lead to insufficient enhancement of key areas or over-processing of non-key areas. However, this invention introduces the concepts of local quality feature analysis and high-interest vertebral segments to achieve semantic-based, non-uniform image enhancement. This can accurately identify poor-quality but anatomically important areas in the image, such as the cervical and lumbar spine, and then use computationally expensive but highly effective super-resolution reconstruction or edge enhancement algorithms for these areas. For similar problems in non-key areas, efficient conventional algorithms are used. This strategy, with a limited computing budget, maximizes the recovery and enhancement of key information on which the final diagnostic decision depends, thus achieving better identification accuracy with higher overall efficiency.

[0116] A processing method for spinal image identification includes the following steps:

[0117] S1. Acquire raw spinal images and generate standardized spinal images through the image preprocessing module;

[0118] S2. Use the vertebral identification and localization module to detect the location and type of vertebrae in standardized spinal images;

[0119] S3. Based on the detection results, generate vertebral labels through the label generation module and overlay them onto the standardized spinal image output;

[0120] S4. Perform quality verification on the images before output: Through preliminary screening, distinguish between images that have passed, are in a critical state, and have failed. Perform dual-path differential processing and consistency assessment on images in a critical state. If the difference exceeds the limit, an early warning will be triggered.

[0121] It should be noted that the calculation formulas and all parameters involved in the calculations in this invention have been dimensionless beforehand. The process of dimensionless processing is well known in the industry and will not be described here.

[0122] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A processing system for spinal image identification, characterized in that, include: The spine image acquisition module is used to acquire raw medical images containing the spine. The image preprocessing module is used to preprocess the original medical images to obtain standardized spinal images; the preprocessing includes: image size normalization, contrast enhancement, noise filtering and image rotation correction; The quality verification and control module is used to perform quality assessment on the standardized spinal images, and take corresponding processing strategies based on the assessment results. It performs consistency verification on the identification results and outputs a quality warning when the identification risk exceeds the threshold. The vertebral identification and localization module is used to detect vertebrae in the standardized spinal images that have passed quality verification, and output the vertebral location information and type label; The identifier generation module is used to generate vertebral identifiers based on the location information and type label. The identifiers include vertebral serial number, bounding box, segment color differentiation, and medical parameter annotations. The labeling output module is used to overlay the vertebral labels onto the standardized spinal image and output a spinal image with labels. The quality verification and control module specifically includes: The preliminary screening subunit is used to calculate the comprehensive quality score of each standardized spinal image based on multiple preset quality indicators, and output three types of results based on the comparison of the comprehensive quality score with preset thresholds: images that have passed the preliminary screening, images that have failed the preliminary screening, and images in a borderline state. The dual-processing subunit inputs the standardized spinal images that have passed the initial screening to the vertebral identification and localization module to obtain the first identification result and the first confidence level. The standardized spinal images of the critical state are input into the vertebral recognition and localization module using two differentiated strategies to obtain a second recognition result, a second confidence level, a third recognition result, and a third confidence level. For images that clearly fail the initial screening, a manual review instruction is triggered; The consistency assessment subunit is used to calculate the difference between the second identification result and the third identification result. The difference is calculated based on the difference in vertebral center coordinates, the deviation in intervertebral space height, the difference in spinal curvature, and the regional weighting coefficient. The early warning unit is used to output quality early warning information when the fusion decision value F calculated based on the comprehensive quality score and the difference value exceeds a preset limit. The early warning information includes the risk of vertebral sequence disorder or local identification deviation.

2. The processing system for spinal image identification according to claim 1, characterized in that, The vertebral identification and localization module is implemented using a deep learning model, which includes a convolutional neural network (CNN), a recurrent neural network (RNN), or a Transformer. The location information output by the deep learning model includes the coordinates of the vertebral bounding box and the coordinates of the center point. The type labels include cervical vertebrae C1-C7, thoracic vertebrae T1-T12, lumbar vertebrae L1-L5, and sacral vertebrae.

3. The processing system for spinal image identification according to claim 1, characterized in that, The difference value of the consistency assessment subunit is calculated using the following formula: Through the formula: The difference value was calculated. ; Among them, the first one is obtained respectively. Difference in center coordinates of vertebrae , No. The height deviation between the intervertebral segment and the adjacent vertebrae Overall curvature of the spine and the first Regional weighting coefficient of vertebrae ; This is a sequence consistency correction function; when there are duplicates or misordered vertebral type label sequences, The value should be between 0.5 and 1.0, otherwise... ; The spinal curvature in the second identification result. The third identification result is the spinal curvature.

4. The processing system for spinal image identification according to claim 3, characterized in that, The differentiation strategies include fine-tuning of preprocessing parameters, switching of model inference modes, or integration of multiple models.

5. The processing system for spinal image identification according to claim 1, characterized in that, The preset quality indicators include image clarity, noise intensity, and vertebral region integrity. The process of calculating the comprehensive quality score in the preliminary screening subunit is as follows: The comprehensive quality score S, normalized to the 0-1 range, is calculated by weighted summation based on image clarity, noise intensity, and vertebral region integrity. The process by which the preliminary screening subunit outputs three types of results based on the comprehensive quality score S is as follows: When the overall quality score is higher than or equal to the first threshold, it is determined that the preliminary screening has been clearly passed; When the overall quality score is lower than or equal to the second threshold, it is determined that the initial screening has been clearly failed. When the overall quality score is between the first threshold and the second threshold, it is determined to be a critical state image; Wherein, the second threshold is less than the first threshold.

6. The processing system for spinal image identification according to claim 5, characterized in that, The formula for calculating the fusion decision value F is: ; in, The difference value calculated for the consistency evaluation subunit. The scaling factor is the normalization coefficient. The quality sensitivity coefficient is, and ≥1, To prevent the division of zero decimals, S represents the overall quality score; For critical state images, when the fusion decision value F > 1, quality warning information is output.

7. The processing system for spinal image identification according to claim 6, characterized in that, The fine-tuning of preprocessing parameters in the differentiation strategy is based on dynamic adjustment of local quality features of critical state images, specifically including: The dual-processing subunit also includes a local quality feature analysis subunit, which is used to identify low-quality regions and high-interest vertebral segments in standardized spinal images and output region quality scores. The first processing path for critical state images is configured as follows: For low-quality regions corresponding to high-interest vertebral segments, super-resolution reconstruction based on generative adversarial networks is used to improve local detail resolution. For low-quality areas that are not of interest, only bilateral filtering is used for basic noise reduction; The second processing path for the same critical state image is configured as follows: For low-quality regions corresponding to high-interest vertebral segments, a multi-scale fusion algorithm is used to enhance edge information; For low-quality regions that are not of interest, noise interference is suppressed by adaptive threshold pruning; Furthermore, the two processing paths employ vertebral recognition models with different initial weights.

Images