Intelligent osteosarcoma image recognition and classification method and system based on image recognition model

Abstract

The invention relates to the technical field of image recognition, in particular to an osteosarcoma image intelligent recognition and classification method and system based on an image recognition model. The method comprises the following steps: collecting a tumor image and carrying out multi-sequence disturbance equilibrium enhancement to generate an enhanced block tumor image, then obtaining a color standardized image through color space transformation remapping, then extracting a tumor region image and carrying out contrast inversion processing, reconstructing a tumor feature image, and finally, carrying out multi-sequence disturbance equilibrium enhancement on the tumor feature image. Based on a preset image recognition model, tumor edge features are recognized, the edge rule degree is determined, preliminary classification is carried out, a benign tumor image is obtained, surrounding metabolite signal features of the benign tumor image are analyzed, a benign reference group is formed through combination, metabolism state comparison is carried out on tumor feature images, and metabolism difference data are generated. And performing category correction according to the metabolic state mapping data so as to identify the osteosarcoma category. According to the invention, intelligent analysis of tumor features is realized, the image processing flow is optimized, and the processing efficiency is improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of image recognition, and in particular to a bone sarcoma image intelligent recognition and classification method and system based on an image recognition model. BACKGROUND

[0002] Existing bone sarcoma image recognition and classification methods often rely on manual analysis and experience-based judgment, resulting in low accuracy and consistency of recognition, especially when dealing with complex image data, subjective factors and operational errors can cause serious misdiagnosis and missed diagnosis, and the differences and complexity between images make it difficult for traditional methods to adapt to diversified clinical needs, and there is a lack of efficient automated processing means, at the same time, the existing image processing technology has shortcomings in feature extraction and image enhancement, which cannot effectively improve the quality of the image, especially in the case of complex background and low contrast, the extraction efficiency of image features is low, which affects the accuracy of subsequent analysis, especially in multi-sequence image processing, how to effectively integrate the advantages of different sequences becomes a problem to be solved, and traditional methods lack systematicness in feature reconstruction and classification, resulting in inaccurate recognition of tumor edge features. SUMMARY

[0003] Therefore, it is necessary to provide a bone sarcoma image intelligent recognition and classification method and system based on an image recognition model to solve at least one of the above technical problems.

[0004] To achieve the above-mentioned purpose, the bone sarcoma image intelligent recognition and classification method based on the image recognition model comprises the following steps: Step S1: collecting a tumor image; performing multi-sequence disturbance equalization enhancement on the tumor image to obtain an enhanced partitioned tumor image; and performing color space transformation and remapping on the enhanced partitioned tumor image to generate a color standardized image; Step S2: extracting a tumor region image from the color standardized image, and performing contrast inversion processing on the tumor region image to obtain an inverted tumor region image; and reconstructing a tumor feature image based on the tumor region image and the inverted tumor region image; Step S3: identifying a tumor edge feature in the tumor feature image based on a preset image recognition model, and determining a tumor edge regularity degree based on the tumor edge feature; and performing preliminary classification and judgment on the tumor feature image based on the tumor edge regularity degree to generate an initial image category; Step S4: obtaining a benign tumor image; analyzing tumor peripheral metabolite signal features based on the benign tumor image, and inferring a normal tumor metabolism state; and combining the normal tumor metabolism state and the benign tumor image as a benign reference group; Step S5: Compare the metabolic state of tumor feature images with the benign reference group to obtain metabolic difference data; map the metabolic state of tumor feature images based on the metabolic difference data to obtain metabolic state mapping data; correct the initial image category based on the metabolic state mapping data to identify the osteosarcoma category.

[0005] This invention acquires tumor images and performs multi-sequence perturbation equalization enhancement, resulting in enhanced segmented tumor images with significantly improved clarity and contrast, providing a high-quality foundation for subsequent analysis. Color space transformation and remapping ensure image consistency under different conditions, and the resulting color-standardized image effectively eliminates color deviations caused by differences in lighting and imaging equipment. The extracted tumor region images undergo contrast inversion processing to further highlight tumor features. The reconstructed tumor feature image integrates the advantages of multiple processing methods, forming a more accurate image representation. Based on a preset image recognition model, tumor edge feature recognition is performed, improving the accuracy and reliability of edge detection. The judgment of edge regularity provides a clear standard for the initial classification of tumors, effectively reducing the influence of human factors on classification judgment. Acquiring benign tumor images and analyzing the surrounding metabolite signal characteristics creates a benign reference group for subsequent metabolic status analysis. This method provides a solid foundation for metabolic state comparison based on a benign reference group. It can accurately identify metabolic differences in tumor feature images, and the generated metabolic difference data provides an important basis for further image analysis. Metabolic state mapping not only improves the accuracy of image category identification but also realizes dynamic correction of the initial image category. The final identified osteosarcoma category has higher accuracy and reliability. The overall method realizes intelligent analysis of tumor features through the application of image recognition technology, optimizes the image processing workflow, improves processing efficiency, reduces subjective interference in traditional methods, and improves the objectivity and systematicness of data analysis. It provides new ideas for the processing and analysis of complex image data, promotes the application depth of intelligent recognition technology in the field of medical imaging, enhances the ability to identify complex lesions such as osteosarcoma, promotes the advancement of imaging analysis technology, and lays the foundation for future research in related fields.

[0006] This invention also provides an intelligent recognition and classification system for osteosarcoma images based on an image recognition model, used to execute the intelligent recognition and classification method for osteosarcoma images based on an image recognition model as described above. The intelligent recognition and classification system for osteosarcoma images based on an image recognition model includes: The image preprocessing module is used to acquire tumor images; perform multi-sequence perturbation equalization enhancement on the tumor images to obtain enhanced block tumor images; and perform color space transformation and remapping on the enhanced block tumor images to generate color-normalized images. The feature reconstruction and analysis module is used to extract the tumor region image based on the color-normalized image, and to perform contrast inversion processing on the tumor region image to obtain the inverted tumor region image; and to reconstruct the tumor feature image based on the tumor region image and the inverted tumor region image. The edge rule classification module is used to identify the edge features of tumors in tumor feature images based on a preset image recognition model, and to determine the degree of regularity of the tumor edges based on the tumor edge features; the tumor feature images are initially classified and judged based on the degree of regularity of the tumor edges to generate an initial image category; The metabolic reference construction module is used to acquire images of benign tumors; analyze the signal characteristics of metabolites around the tumor based on the images of benign tumors, and infer the normal metabolic state of the tumor; and combine the normal metabolic state of the tumor and the images of benign tumors into a benign reference group. The metabolic correction and identification module is used to compare the metabolic state of tumor feature images with a benign reference group to obtain metabolic difference data; to map the metabolic state of tumor feature images based on the metabolic difference data to obtain metabolic state mapping data; and to perform category correction on the initial image category based on the metabolic state mapping data to identify the osteosarcoma category.

[0007] This invention utilizes an image preprocessing module to enhance acquired tumor images through multi-sequence perturbation equalization, resulting in enhanced, segmented tumor images with significantly improved detail and contrast. Color space transformation and remapping ensures image consistency under different imaging conditions, and the resulting color-standardized image effectively eliminates the influence of lighting differences between devices. A feature reconstruction and analysis module extracts tumor region images and performs contrast inversion processing, resulting in inverted tumor region images that effectively highlight key features. The reconstructed tumor feature image integrates the advantages of multiple processing methods, enhancing feature expressiveness. An edge rule classification module identifies edge features in the tumor feature image based on a preset image recognition model, significantly improving recognition accuracy. The determination of edge regularity provides a scientific basis for preliminary classification, reducing the influence of subjective factors. A metabolic reference construction module acquires benign tumor images and analyzes the surrounding metabolite signal characteristics, forming a benign reference group that provides a strong foundation for subsequent analysis. The inference of metabolic state can effectively capture normal... The metabolic feature and metabolic correction identification module uses a benign reference group to compare the metabolic state of tumor feature images. The generated metabolic difference data provides an important basis for further analysis. The metabolic state mapping based on the metabolic difference data not only improves the accuracy of image classification but also realizes dynamic correction of the initial category. The final identified osteosarcoma category has higher accuracy and reliability. The overall system achieves efficient processing of complex image data through the application of intelligent recognition technology, optimizes the image analysis process, improves data processing efficiency, reduces subjective interference in traditional methods, and enhances the objectivity and systematicness of data analysis. It provides new ideas for future research in related fields, promotes the in-depth application of intelligent recognition technology in the field of medical imaging, improves the ability to identify complex lesions such as osteosarcoma, promotes the advancement of imaging analysis technology, lays a solid foundation for the application of large-scale data processing scenarios, and promotes the combination of storage-computing technology and intelligent recognition technology, forming an efficient and accurate data processing system. Attached Figure Description

[0008] Figure 1 This is a flowchart illustrating the steps of an intelligent identification and classification method for osteosarcoma images based on an image recognition model. Figure 2 This is a detailed flowchart illustrating the implementation steps of step S1. The realization of the objective, functional features and advantages of the present invention will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0009] The technical method of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without inventive effort are within the scope of protection of the present invention.

[0010] Furthermore, the accompanying drawings are merely illustrative of the invention and are not necessarily drawn to scale. The same reference numerals in the drawings denote the same or similar parts, and therefore repeated descriptions of them will be omitted. Some block diagrams shown in the drawings are functional entities and do not necessarily correspond to physically or logically independent entities. These functional entities can be implemented in software, in one or more hardware modules or integrated circuits, or in different network and / or processor methods and / or microcontroller methods.

[0011] It should be understood that although the terms "first," "second," etc., may be used herein to describe various units, these units should not be limited by these terms. These terms are used merely to distinguish one unit from another. For example, without departing from the scope of the exemplary embodiments, a first unit may be referred to as a second unit, and similarly, a second unit may be referred to as a first unit. The term "and / or" as used herein includes any and all combinations of one or more of the associated listed items.

[0012] To achieve the above objectives, please refer to Figures 1-2 An intelligent identification and classification method for osteosarcoma images based on an image recognition model includes the following steps: Step S1: Acquire tumor images; perform multi-sequence perturbation equalization enhancement on the tumor images to obtain enhanced block tumor images; perform color space transformation and remapping based on the enhanced block tumor images to generate color-normalized images; Step S2: Extract the tumor region image from the color-normalized image, and perform contrast inversion processing on the tumor region image to obtain the inverted tumor region image; reconstruct the tumor feature image based on the tumor region image and the inverted tumor region image; Step S3: Identify the tumor edge features in the tumor feature image based on the preset image recognition model, and determine the regularity of the tumor edge based on the tumor edge features; perform preliminary classification judgment on the tumor feature image based on the regularity of the tumor edge to generate an initial image category; Step S4: Obtain images of benign tumors; analyze the signal characteristics of metabolites around the tumor based on the images of benign tumors, and infer the normal metabolic state of the tumor; combine the normal metabolic state of the tumor and the images of benign tumors into a benign reference group; Step S5: Compare the metabolic state of tumor feature images with the benign reference group to obtain metabolic difference data; map the metabolic state of tumor feature images based on the metabolic difference data to obtain metabolic state mapping data; correct the initial image category based on the metabolic state mapping data to identify the osteosarcoma category.

[0013] This invention acquires tumor images and performs multi-sequence perturbation equalization enhancement, resulting in enhanced segmented tumor images with significantly improved clarity and contrast, providing a high-quality foundation for subsequent analysis. Color space transformation and remapping ensure image consistency under different conditions, and the resulting color-standardized image effectively eliminates color deviations caused by differences in lighting and imaging equipment. The extracted tumor region images undergo contrast inversion processing to further highlight tumor features. The reconstructed tumor feature image integrates the advantages of multiple processing methods, forming a more accurate image representation. Based on a preset image recognition model, tumor edge feature recognition is performed, improving the accuracy and reliability of edge detection. The judgment of edge regularity provides a clear standard for the initial classification of tumors, effectively reducing the influence of human factors on classification judgment. Acquiring benign tumor images and analyzing the surrounding metabolite signal characteristics creates a benign reference group for subsequent metabolic status analysis. This method provides a solid foundation for metabolic state comparison based on a benign reference group. It can accurately identify metabolic differences in tumor feature images, and the generated metabolic difference data provides an important basis for further image analysis. Metabolic state mapping not only improves the accuracy of image category identification but also realizes dynamic correction of the initial image category. The final identified osteosarcoma category has higher accuracy and reliability. The overall method realizes intelligent analysis of tumor features through the application of image recognition technology, optimizes the image processing workflow, improves processing efficiency, reduces subjective interference in traditional methods, and improves the objectivity and systematicness of data analysis. It provides new ideas for the processing and analysis of complex image data, promotes the application depth of intelligent recognition technology in the field of medical imaging, enhances the ability to identify complex lesions such as osteosarcoma, promotes the advancement of imaging analysis technology, and lays the foundation for future research in related fields.

[0014] In this embodiment of the invention, the intelligent recognition and classification method for osteosarcoma images based on an image recognition model includes the following steps: Step S1: Acquire tumor images; perform multi-sequence perturbation equalization enhancement on the tumor images to obtain enhanced block tumor images; perform color space transformation and remapping based on the enhanced block tumor images to generate color-normalized images; In this embodiment, the Philips Achieva 3.0T MRI system was used to acquire images of the tumor in the patient's affected area. During acquisition, three imaging sequences were used: T1-weighted, T2-weighted, and enhanced T1 (T1-CE). The slice thickness was set to 5 mm, the slice interval to 0.5 mm, the field of view to 256 × 256 mm, and the acquired image size to 512 × 512 pixels. The image format was DICOM. All acquired images were uniformly decoded and batch preprocessed using the open-source image preprocessing tool SimpleITK. After decoding, the images underwent multi-sequence perturbation equalization enhancement. In the enhancement stage, each image sequence was first divided into blocks, with a block size of 32 × 32 pixels. Each image block underwent local contrast-limited histogram equalization (CLAHE) with a contrast limit value of 2.0. The limit value is set to 40. After enhancement, the image is reconstructed by image stitching. The enhanced image then undergoes color space transformation. The cv2.cvtColor function in OpenCV is used to convert the RGB image to the CIELAB color space, where the L channel is used for luminance remapping. The contrast enhancement effect is adjusted through exponential transformation, and the exponential function coefficients are... Setting it to 0.85, the remapped CIELAB image is then converted back to RGB space, resulting in a color-normalized image. The average standard deviation of the color distribution across different images is controlled within 5.

[0015] Step S2: Extract the tumor region image from the color-normalized image, and perform contrast inversion processing on the tumor region image to obtain the inverted tumor region image; reconstruct the tumor feature image based on the tumor region image and the inverted tumor region image; In this embodiment, after obtaining the color-normalized image, the depth-based instance segmentation network YOLACT (You Only Look At CoefficienTs) is invoked to accurately extract the tumor region in the image. The backbone of this model is ResNet-101, and it incorporates an FPN structure for multi-scale region localization. During training, IoU (Intersection over Union) is used. Union was used as the detection accuracy indicator, and the confidence threshold was set to 0.7. Only the main target mask region of the tumor area was retained. The extracted tumor area image was then subjected to pixel-level inversion processing. The gray value g(x,y) of each pixel in the grayscale image became 255-g(x,y) after inversion. The inverted tumor area image and the original tumor area image were used as input to a U-Net network with residual connection structure for image reconstruction. The network structure contains four encoder layers and four decoder layers. The encoder part uses 3×3 convolution kernels for feature extraction with a stride of 1. The decoder part uses deconvolution to achieve feature upsampling. The input image size is 256×256 pixels. During the reconstruction process, the loss function adopts a weighted combination of structural similarity loss (SSIM) and mean squared error loss (MSE). The loss weights are set to SSIM:0.7 and MSE:0.3. Finally, the fused tumor feature image is output.

[0016] Step S3: Identify the tumor edge features in the tumor feature image based on the preset image recognition model, and determine the regularity of the tumor edge based on the tumor edge features; perform preliminary classification judgment on the tumor feature image based on the regularity of the tumor edge to generate an initial image category; In this embodiment, the tumor feature image is input into a trained image recognition model for feature analysis. The image recognition model adopts a Deeplabv3+ structure based on an edge-guided mechanism, with Xception as the backbone network. The dilation rate of the dilated convolution is set to 6, 12, and 18 for three-layer feature extraction. The extracted image edge feature map is then sent to the edge regularity judgment module. This module uses Freeman chain code to encode the edge contour lines and uses Zernike moment features to quantify the symmetry, smoothness, and closure of the contour graphics. Edge contours with a symmetry coefficient less than 0.1 and a closure greater than 95% are labeled as regular edges, and those with other values ​​are labeled as irregular edges. The edge regularity label is used as a priori feature for image category judgment and input into a support vector machine (SVM) classifier for preliminary classification. The classification label output is the initial image category, with 0 for benign categories and 1 for malignant osteosarcoma categories.

[0017] Step S4: Obtain images of benign tumors; analyze the signal characteristics of metabolites around the tumor based on the images of benign tumors, and infer the normal metabolic state of the tumor; combine the normal metabolic state of the tumor and the images of benign tumors into a benign reference group; In this embodiment, a benign tumor image dataset constructed in the database is used as the benign reference sample data, with 300 samples. All images are pre-processed to a size of 256×256 pixels, in PNG format, and in RGB color space. The benign image samples are then paired with proton magnetic resonance spectroscopy (1H-MRS) data. Metabolite signals are extracted from the edge region of each sample. The Cho (choline), Cr (creatine), and NAA (N-acetylaspartate) signal peaks are fitted using the Lorentzian curve fitting method. The Cho / Cr and NAA / Cr ratios are calculated to construct a metabolite signal feature vector. The metabolic stability threshold range is set as Cho / Cr≤1.2 and NAA / Cr≥1.6. If the range is met, it is considered a normal metabolic state. Each benign image and its metabolic feature vector are input into a deep encoder for feature fusion. The deep encoder structure is a multi-scale convolutional embedding module, which finally generates a benign reference group feature vector with a dimension of 128. All reference vectors are aggregated through mean pooling to generate benign metabolic baseline features.

[0018] Step S5: Compare the metabolic state of tumor feature images with the benign reference group to obtain metabolic difference data; map the metabolic state of tumor feature images based on the metabolic difference data to obtain metabolic state mapping data; correct the initial image category based on the metabolic state mapping data to identify the osteosarcoma category.

[0019] In this embodiment, the aforementioned tumor feature image and benign metabolic baseline features are input into the metabolic state comparison module. The metabolic state comparison module adopts a Siamese network structure. The main branch extracts the structural and metabolic coupling features of the tumor image, and the auxiliary branch extracts the metabolic features of the benign reference. Cosine similarity is used as the difference measure function to output a metabolic difference vector. The difference vector is mapped to the original image space, and the spatial distribution map is reconstructed through an invertible residual network to generate a metabolic difference map with a dimension of 256×256. The metabolic difference map and the tumor feature image are fused at the channel level. After fusion, the image is input into the image attention mapping module for metabolic state space projection. After projection, a metabolic state mapping map is obtained. This mapping map and the initial image category label are input into the image category correction module. The category correction module is a dual-input dual-output Transformer structure. The initial classification label is corrected using a graph attention mechanism, and the final image category label is output. The label value is the osteosarcoma identification result. The final identification result is output as a structured data table and used for subsequent statistical analysis.

[0020] Preferably, step S1 includes the following steps: Step S11: Acquire tumor images; perform sequence splicing and fusion on the tumor images to obtain a multi-sequence tumor fusion image, wherein the number of spliced ​​sequences is limited to no less than 3 sets; inject perturbation information into the multi-sequence tumor fusion image data to generate a perturbation-enhanced tumor image; Step S12: The perturbation-enhanced tumor image is locally equalized and divided into blocks to generate a block equalized tumor image, wherein the size of each block is set to 32×32 pixels; the block equalized tumor image is subjected to multi-dimensional scale detail enhancement processing to obtain an enhanced block tumor image, wherein the enhancement scale level is set to 3-5 layers. Step S13: Perform edge-preserving smoothing processing on the enhanced block tumor image to generate an edge-enhanced smoothed tumor image; perform color space remapping on the edge-enhanced smoothed tumor image to obtain a color space-mapped tumor image; Step S14: Perform hue normalization processing on the tumor image based on color space mapping and perform pseudo-color encoding to obtain a pseudo-color tumor image, wherein the hue mapping range is set to 0°-270°. Step S15: Perform color layer resampling based on the pseudo-color tumor image to generate a color-normalized image.

[0021] In this embodiment, a Philips Achieva 3.0T MRI scanner is used to acquire images of the target area. The scanning parameters are set to three sequences: T1-weighted, T2-weighted, and T1-enhanced contrast sequence. Each sequence acquires at least five image layers. The image resolution is set to 512×512 pixels, the layer thickness is 4mm, the interlayer spacing is 0.6mm, and the image acquisition format is DICOM. After image acquisition, ITK (Insight Segmentation and...) is used... The RegistrationToolkit tool performs spatial alignment and fusion processing on three sets of sequential images. Affine transformation and trilinear interpolation are used to overlay the three images onto a unified spatial coordinate system, generating a multi-sequence fused tumor image. Gaussian perturbation is introduced into the fused image to achieve perturbation enhancement. The perturbation process uses a Gaussian distribution with a mean of 0 and a standard deviation of 0.05 to randomly inject perturbations into pixel values, followed by pixel intensity normalization to limit the pixel value range to [0,1]. After perturbation enhancement, the perturbation-enhanced tumor image is output. The OpenCV library's image block division function performs local equalization block division on the perturbation-enhanced tumor image, dividing the image into 32×32 pixel grid blocks. Histogram equalization is performed within each block, and OpenCV's equalizeHist method is used to stretch the brightness value of each block individually, adjusting the local grayscale distribution to an approximately uniform distribution, thus generating a block-equalized tumor image. Finally, multi-dimensional scale detail enhancement is performed on this image, using a Laplacian pyramid structure to construct a multi-scale pyramid hierarchy. The pyramid structure is set to 5 layers. Local high-frequency residual maps are calculated for each layer, and high-frequency details are cumulatively added back to the original image layer by layer. The convolution kernel size for each high-frequency enhancement is 3×3, and the enhancement weights are progressively decreased to 0.8, 0.6, 0.4, 0.2, and 0.1, resulting in enhanced segmented tumor images. Edge-preserving smoothing is then performed on these images. A bilateral filter-based algorithm is used to smooth and denoise the image, with the spatial distance Gaussian kernel σs set to 15, the color distance Gaussian kernel σr set to 0.1, and the convolution window size set to 5×5. This preserves image edges while eliminating noise and smoothing low-frequency background areas, outputting an edge-enhanced and smoothed tumor image. This image is then input into a color space remapping module. The color space remapping uses a conversion process from RGB to CIELAB color space, implemented using the `skimage.color.rgb2lab` function in Python to transform the image from RGB space to the CIELAB three-channel space. During the remapping process, Gamma nonlinear correction is applied to the L channel, with the Gamma coefficient set to 0.9. Standardize the a and b* channels to the [-1,1] interval. After completing the color space mapping, output the color space mapped tumor image. Process the color space mapped tumor image using a hue normalization model. This normalization model normalizes the dominant hue in different images based on color cluster centers. Use the K-means clustering algorithm to cluster the a and b channels, setting the number of clusters to 8. Each cluster is matched to a preset reference hue space, which is derived from the average color histogram in the osteosarcoma medical image training set. Then, perform pseudo-color encoding to assign a specified hue code to different brightness areas in the image. Use the OpenCV applyColorMap function to perform pseudo-color mapping. The mapping adopts linear allocation of the H channel in the HSV color space, with the hue range set between 0° and 270°, based on pixel grayscale. The image is mapped to different hues, ultimately outputting a pseudo-color tumor image. This pseudo-color tumor image is then input into a color grading resampling module for color standardization. The color grading module uses a fixed number of color clusters to gradate the image, dividing each image into 9 color level intervals based on the L channel brightness. The brightness intervals are distributed as [0-30], [30-60], ..., [240-270]. For each brightness-graded image block, PCA (Principal Component Analysis) is used to project and resample the color distribution direction. The color information is projected onto a two-dimensional principal component space and then reconstructed using KNN interpolation. After reconstruction, the image color space is remapped back to RGB format, outputting a color-standardized image. The standard deviation of the image color is controlled within 8 for each of the three channels, and the Bhattacharyya distance between the color histogram and the standard training set samples is controlled within 0.05.

[0022] Preferably, step S2, which involves extracting the tumor region image from the color-normalized image and performing contrast inversion processing on the tumor region image, includes: Identify tumor region boundaries based on color-standardized images; The tumor region boundary is expanded inward to obtain the expanded tumor region. The tumor region image is extracted from the color-normalized image by expanding the tumor region; The contrast inversion process is performed on the tumor region image to obtain the initial inverted tumor region image; Shape correction processing is performed on the inverted tumor region image to obtain the inverted tumor region image.

[0023] In this embodiment, in the intelligent recognition and classification method for osteosarcoma images based on image recognition models, after the color-normalized image is acquired, it is input into an osteosarcoma region detection model based on deep semantic segmentation. This model uses the SegFormer-B2 version, with the backbone network employing a MixVision Transformer structure to extract global contextual features. The input image size is set to 512×512 pixels. Image preprocessing includes brightness normalization and boundary filling. The model is trained using a combination of cross-entropy loss and Dice loss. Softmax normalization is performed on the output to obtain a pixel-level segmentation probability map. A segmentation threshold of 0.5 is set, and connected component analysis is performed on regions above the threshold. The main connected component with the largest area is retained as the tumor region boundary contour. The output is a tumor region boundary mask. The extracted tumor boundary is a two-dimensional closed curve, and a morphological erosion algorithm is used to refine this boundary. A pixel-level expansion operation is performed inward. The erosion kernel size is set to a 5×5 matrix, the structuring element type is set to an elliptical kernel, and the number of erosion iterations is set to 2. The boundary shape after erosion maintains the morphological characteristics of the tumor while shrinking the boundary width inward by approximately 8-10 pixels, generating a new expanded tumor region mask. The mask image is multiplied pixel-by-pixel with the original color-normalized image, and the original image content within the mask region is extracted to form the tumor region image. The image size is adaptively cropped by the mask boundary, and the boundary zero-padding area is filled using a mirror fill algorithm. The tumor region image is then input into the contrast inversion module, where the image undergoes channel-by-channel inversion processing. The sequence involves three channels: R, G, and B. The pixel value p(i,j) for each channel is processed using the formula 255 - p(i,j). All pixel values ​​are operated on within an 8-bit integer range. After image inversion, the overall brightness distribution undergoes a symmetrical transformation, with low-brightness areas becoming high-brightness areas and prominent contours being enhanced. The inverted image is named the initial inverted tumor region image. This image is then input into the shape correction module for further processing. The shape correction module performs an affine transformation based on the boundary shape of the inverted image, first using the `findContours` function from the OpenCV library to extract the main edges of the inverted image. For the contour line, the minimum bounding rectangle method is used to fit and calculate its aspect ratio and tilt angle. If the contour tilt angle is greater than 10°, an affine rotation transformation is performed. The rotation center is set to the geometric center of the image, and the rotation angle is set to -θ, where θ is the original tilt angle. The rotation transformation matrix is ​​constructed using the getRotationMatrix2D function, and the image is then subjected to warpAffine interpolation mapping. Bilinear interpolation is used, and the edge filling method uses a constant fill value of 0. After shape correction, the inverted tumor region image is output for subsequent feature fusion and image reconstruction steps.

[0024] Preferably, the reconstruction of the tumor feature image based on the tumor region image and the inverted tumor region image in step S2 includes: Affine registration is performed on the tumor region image and the reversed tumor region image, where the affine matrix is ​​limited to the rotation angle [-10°, +10°] and the scaling ratio [0.9, 1.1] to obtain the initial aligned image; Multi-channel residual enhancement is performed on the initial aligned image to generate a residual fused image; Guided edge filtering is applied to the residual fusion image to obtain a boundary optimized image, wherein the guide kernel size is set to 5×5 and the edge threshold range is set to [20, 50]. The boundary optimization image is subjected to morphological semantic compression mapping to obtain the tumor feature image.

[0025] In this embodiment, during the image feature construction process for the intelligent recognition and classification task of osteosarcoma images driven by the image recognition model, affine registration is first performed on the extracted tumor region image and the inverted tumor region image after shape correction. The affine registration process uses the `estimateAffinePartial2D` function in OpenCV. The input image pair consists of the original and inverted tumor region images. Corresponding key points are extracted and matched using the ORB (Oriented FAST and Rotated BRIEF) algorithm, with the number of key points controlled to around 500. The matching method uses Hamming distance for brute-force matching. To ensure the stability and semantic consistency of the transformation, the registration matrix is ​​set to a 2×3 affine matrix. The rotation angle parameter of the affine transformation is limited to [-10 degrees, +10 degrees], and the scaling factor is limited to the range of [0.9, 1.1]. Boundary clipping is performed on the affine parameters in the estimated matrix, and affine values ​​exceeding the set range are set as boundary values. The affine transformation execution function is warpAffine, and bilinear interpolation is used. The final output size is consistent with the input image, generating an initial aligned image pair, representing the registered tumor image and the registered inverted image, respectively. The initial aligned images are input to the multi-channel residual enhancement module. First, a channel-level stacking operation is performed on the two images to form a six-channel input tensor, which is then fed into an enhancement network containing deep residual units for processing. The network structure is based on the ResNet-34 backbone. The model incorporates spatial and channel attention mechanisms in each residual block. The channel attention mechanism uses the Squeeze-and-Excitation (SE) module to adaptively weight the feature tensor along the channel dimension. The spatial attention mechanism uses a combination of 3×3 convolution and sigmoid to generate a spatial mask. The input size of the enhanced network is 256×256×6, and the output feature map is 256×256×3. The output result is the fused residual image, which is named the residual fused image. The residual fused image is then input into the guided edge filtering module, which employs a weighted smoothing filtering algorithm based on the guided image. The image itself is used as the guided image, and a weighted average value is calculated within a 5×5 neighborhood window for each pixel. The weights are calculated using a combination function of color distance and spatial distance, with the color weight term σr set to 0.15. The spatial weight term σs is set to 2, the filter kernel size is fixed at 5×5, the edge preservation strength is controlled by the edge threshold, and pixels with edge response values ​​in the range of [20, 50] are set to retain details, while the rest are smoothed. The edge response map is calculated using the Sobel gradient operator, and the final output is a boundary-optimized image with more continuous and smooth boundary contours and significantly reduced background noise. The boundary-optimized image is input to the morphological semantic compression mapping module for image feature reconstruction. This module uses a semantic compression structure based on spatial pyramid pooling (SPP) to encode the image in a multi-level structure. The SPP module is divided into 1 The image is divided into three pooling branches: ×1, 2×2, and 4×4. Each branch performs average pooling, generating semantic vector representations at different receptive field scales. The results from all branches are concatenated and fed into a convolutional layer to map to a uniform 128-dimensional semantic vector. This vector is then reconstructed through a fully connected layer to an image size of 256×256×1. Bilinear interpolation is used to restore the original image size. After compression, the intensity of unstructured regions in the image is reduced, while structural regions retain high response. The final output image is named the tumor feature image, which serves as key feature input data in subsequent edge recognition, classification, and metabolic mapping steps.

[0026] Preferably, step S3, which involves identifying tumor edge features in a tumor feature image based on a preset image recognition model and determining the regularity of the tumor edge based on these features, includes: Extract the tumor edge image from the tumor feature image based on a preset image recognition model; Remove false edges from the tumor edge image and extract the true contour of the tumor; Based on the true contour of the tumor, the edge features of the tumor are identified; The overall morphological analysis was performed based on the characteristics of the tumor margins, and the results were quantified into overall morphological characteristic indicators. Map the outer contour lines of the tumor to accurately depict its true outline; Based on the actual contour of the tumor, the outer contour lines are located by a sequence of equally spaced points to generate equally spaced contour points. Constructing a local neighborhood of contour points based on equally spaced contour points; Local undulation gradients of each subdomain are analyzed based on the local neighborhood of contour points. The regularity of the tumor margin is determined by assessing the overall morphological characteristics and the local undulation gradient of each sub-domain.

[0027] In this embodiment, in the intelligent recognition and classification method for osteosarcoma images based on image recognition models, edge recognition and regularity evaluation operations are performed on the input tumor feature image. First, the tumor feature image is input into the deep edge detection network EdgeNet based on the Canny attention fusion structure. This model uses ResNet-50 as the backbone extractor and fuses an edge guidance branch. The guidance branch inserts 1×1 convolutions in the first two layers of the network to generate edge attention weights. The model output is a single-channel tumor edge image with pixel values ​​ranging from [0,1], representing the edge probability. The probability map is hard segmented with a binarization threshold of 0.35 to generate a preliminary tumor edge image. The preliminary edge image is then subjected to pseudo-edge removal. Morphological opening operation and small region area filtering are used together to remove interfering edges. The opening operation uses an elliptical structuring element with a kernel size of 3×3 and performs one erosion and dilation operation. Subsequently, connected component analysis is performed on the remaining edge regions, retaining areas greater than 0.5% of the edge-closed area was removed, and the remaining edge area was used as the true contour image of the tumor. The true contour line was represented by white pixel values ​​(255). Edge feature extraction was performed on the true contour image of the tumor. First, the outer contour curve was extracted using the edge contour tracking function findContours. Then, the curvature, edge length, average edge angle, and edge closure of each contour were calculated. The edge length was obtained by calculating the sum of the Euclidean distances between contour points. The edge angle was calculated by averaging the angles formed by vectors composed of three adjacent contour points. The curvature was obtained by fitting a quadratic polynomial to the contour. The residual standard deviation is calculated, and the edge closure is derived by inversely using the ratio of the Euclidean distance between the start and end points to the boundary length. The above four indicators are normalized and encoded, and combined into an overall morphological feature index vector. The outer contour lines of the tumor's true outline are uniformly sampled, with a fixed number of sampling points of 360. An equidistant contour point set is generated using a uniform arc length distribution method. Uniform arc length sampling is achieved through curve integration to ensure the geometric consistency of the contour points in morphological distribution. The equidistant contour point sequence is a two-dimensional coordinate sequence, with each point representing a key structural node on the contour boundary in the image. A local neighborhood structure is constructed around the equidistant contour points, defined as... A seven-point structural unit, consisting of three points extending to the left and right from the current point, is constructed. A sliding window moves across the entire contour point sequence to generate 360 ​​sets of local neighborhood structures. Each neighborhood structure contains the coordinate information of the point, which is used for subsequent local undulation gradient analysis. Local undulation gradient analysis is performed on the local neighborhood of the contour point. The undulation gradient is defined as the sum of the vertical height differences between the current point and its left and right symmetrical points. It is calculated using the three-point symmetric difference method, that is, the slope of the line connecting the previous and next point pairs is calculated with the current point as the center point. Furthermore, the local curvature response of the point is estimated by the numerical first derivative function. The undulation gradient of each point is recorded in a vector to form the local gradient. The distribution map shows that fluctuations exceeding 1% of the image diagonal length are identified as significant abrupt changes. Overall morphological features and local gradient maps are combined for evaluation and input into the classification rule assessment module. This module constructs a regularity discrimination logic based on a decision tree model. Irregularity indicators include larger edge lengths and greater average angle deviations, more local gradient abrupt changes, and lower edge closure. The model is trained on 1280 osteosarcoma image feature samples using five-fold cross-validation. Ultimately, the model classifies the image regularity as regular or irregular and outputs a regularity score, which serves as an important prior basis for subsequent preliminary image classification and input into the next module for further processing.

[0028] Of particular importance is the assessment of tumor regularity using overall morphological characteristics and local undulation gradients in each sub-region, including: The overall morphological feature index is coupled with the local undulation gradient of each subdomain to obtain the fused morphological undulation feature. Perform edge fitting consistency retrieval on the fusion morphological fluctuation features to generate fitting consistency response data; Based on the fitted consistent response data, a rule degree level mapping is performed to obtain rule level mapping data; Based on the rule level mapping data, the contour stability of the tumor feature image is corrected to generate edge rule stability image data; The edge regularity stability image data is threshold-classified based on a preset regularity threshold to determine the degree of regularity of the tumor edge.

[0029] In this embodiment, a normalized feature coupling operation is performed on the overall morphological feature indicators obtained from previous calculations and the local undulation gradients of each subdomain. First, the numerical dimensions of each dimension in the overall morphological feature indicators are standardized to the [0,1] interval. The maximum and minimum normalization methods are used to normalize the four indicators of curvature, average edge angle, edge length, and closure, respectively. The maximum and minimum values ​​are obtained based on the statistical distribution of the entire training set. The local undulation gradient vectors are normalized using the Z-score normalization method, and the mean and standard deviation are also derived from the statistical distribution of each location point in the training set. Subsequently, the normalized overall indicators and local undulation gradients are coupled. Gradient vectors are concatenated along the feature dimensions to form a unified fused morphological undulation feature vector. The total dimension of the vector is 364, with the first 4 dimensions representing global indices and the remaining 360 dimensions representing the undulation values ​​at equally spaced points. This fused morphological undulation feature is input into the edge fitting consistency retrieval module. The module loads five predefined standard edge morphology templates: ellipse, circle, regular pentagon, wavy boundary, and jagged boundary. Each template is represented by a morphological contour vector generated from 1000 real samples through linearly normalized edge averaging. The Dynamic Time Warping (DTW) algorithm is then used. Time Warping (DTW) calculates the fit consistency between the fused features and each template. Euclidean distance is used as the distance metric in DTW calculations, and the sequence length is uniformly padded to 360. The final output consists of five fit consistency response values, each corresponding to an edge template. Lower values ​​indicate a higher fit. The five response values ​​are sorted in ascending order, and the template with the smallest fit value is selected as the reference template. The corresponding consistency value is recorded as the fit response output, forming the fit consistency response data. This data is then input into a regularity level mapping module. This module has five preset regularity mapping intervals, named as regular, sub-regular, slightly abrupt boundary changes, moderate changes, and severe changes. The mapping interval thresholds are set to [0–0.15], [0.15–0.25], [0.25–0.35], [0.35–0.5], and [0.5–1]. The response values ​​are mapped to specific levels. If the fit consistency response value is 0...The 18 rules correspond to the "sub-rules" of the rule level. The level label is added to the image space as a one-dimensional rule level matrix as rule level mapping data. This label is subsequently used in the adjustment and stability correction of the contour image. The rule level mapping data is used as the control condition input to the contour stability correction module to perform structural contour reconstruction and boundary compensation on the tumor feature image. Contour reconstruction uses the edge response map for gradient backpropagation. Through depth-guided deconvolution, the rule level control parameters are mapped to image-level structural morphological constraints and superimposed on the main channel of the feature map for edge reconstruction. Edge correction uses a directional filter bank to perform response convolution along the contour direction. The filter size is 7×7. The edge stability is generated as edge rule stability image data based on the point-to-point gradient stability response statistics. The image size is the same as the original. Figure 1 The single-channel grayscale value ranges from 0 to 255, with higher values ​​indicating higher edge regularity. The edge regularity stability image data is input into the threshold grading module, which loads preset regularity stability thresholds. The low stability threshold is set to 90, and the high stability threshold is set to 180. Each pixel value in the stable image is divided into three categories according to the threshold: values ​​less than 90 are marked as low-regularity regions, values ​​greater than 180 are marked as high-regularity regions, and the rest are medium-regularity regions. The proportion of each region within the entire tumor boundary is counted. If the proportion of high-regularity regions exceeds 70% and low-regularity regions are less than 10%, the overall tumor edge is judged as regular. If the proportion of low-regularity regions exceeds 30%, it is judged as irregular edge, and the rest are judged as medium-regularity tumor edges. Finally, the tumor edge regularity determination label is output and bound to the tumor image metadata structure for subsequent classification model calls and processing.

[0030] Preferably, the preliminary classification and judgment of the tumor feature image based on the regularity of the tumor edge in step S3 includes: The regularity data of tumor edges are expanded in feature dimensions to generate a multidimensional expression vector of regularity. The regularity-level multidimensional representation vector is subjected to feature clustering to obtain regular feature cluster labels; Identify rule structure anomalies in the rule feature cluster labels and mark them as rule structure anomaly labels; Graph structure encoding is performed on the regular feature cluster labels and structural anomaly labels to generate a comprehensive regular morphological encoding. The regular morphology is comprehensively encoded, then decoded as a feature, and mapped to a regular morphology map of the tumor. Based on the regular morphology atlas of tumors, a preliminary classification judgment is made on the tumor feature images to generate initial image categories.

[0031] In this embodiment, the previously calculated tumor edge regularity data undergoes feature dimension expansion processing. First, the original scalar value of the regularity and the response value of each pixel in the edge regularity stability image are loaded. Statistical descriptive features such as mean, standard deviation, maximum, minimum, and regional distribution ratio are calculated through local statistical operations. Simultaneously, the regularity level labels on the contour lines are mapped to one-hot encoding, with five levels of regularity corresponding to five-dimensional one-hot vectors, which are then appended to the overall feature vector. Based on this, a multi-scale window extraction operation is performed, with window sizes set to 32×32, 64×64, and 128×128, and window sliding steps set to 8, 16, and 32 pixels respectively. The mean of the regularity response features and the consistency value of the edge direction within each window region are statistically analyzed. Finally, all dimensions are concatenated to form a multi-dimensional regularity expression vector. The generated feature vector has 512 dimensions, with each image sample corresponding to one rule. The degree-based multidimensional vector representation structure is used. The degree-based multidimensional representation vector is input into the K-means++ clustering model for unsupervised clustering. The number of clusters is set to 4. Initial centroids are selected using a density-weighted maximum margin strategy. The number of iterations is set to 500, and the stopping condition is that the centroid change is less than 1e-5. Finally, regular feature cluster labels are generated, with label values ​​of four categories: {C0, C1, C2, C3}. The label numbers are sorted from low to high according to the average degree of regularity of the centroids. All samples are assigned to a category, forming a preliminary structural feature cluster mapping map. The output is the category number to which each image sample belongs. A structural consistency evaluation operation is performed on the regular feature cluster labels. By calculating the pairwise cosine similarity between vectors in each class, the mean and standard deviation are used to form an intra-class compactness index. If the similarity between a sample vector and the centroid vector in a certain class is lower than a set threshold (threshold set to 0.65) and the similarity between the centroids of other classes is greater than or equal to 0, the clustering is considered complete.If the value is 70, the sample is determined to be a structural outlier. All structural outliers are uniformly assigned anomaly label E, which records their original category number and deviation direction, forming a regular structural outlier label data structure. The regular feature cluster labels and regular structural outlier labels are input into the graph structure encoding module. This module constructs a graph neural network model, treating each sample as a node in the graph. The node feature is a regularity vector. Fully connected edges are established between all nodes within each category, with edge weights being cosine similarity values. Outlier nodes also form cross-cluster connections with their deviation edges from the target category. The graph structure uses a GCN (Graph Convolutional Network) model for two rounds of graph convolution operations, with a node embedding dimension of 256 and ReLU activation function. The output is a regular morphological graph embedding encoding for each node, named the regular morphological comprehensive encoding, with a uniform dimension of 128. The regular morphological comprehensive encoding is input into the decoding module for feature reconstruction. The decoding structure uses a fully connected multilayer perceptron (MLP). The Perceptron architecture has a 128-dimensional input and two hidden layers with dimensions of 256 and 512 respectively. The output is an image spatial morphology map, 256×256 pixels, a single-channel grayscale image. Pixel values ​​represent the regularity response level of local regions within the graphic structure. The map exhibits a distribution of regular center aggregation and prominent abnormal edges, named the tumor regular morphology map. The tumor regular morphology map and the tumor feature image are jointly matched and mapped. The input is a discriminative initial classification module, a lightweight convolutional neural network with a two-branch structure. The main branch extracts the spatial structure features of the tumor feature image, and the auxiliary branch extracts regional features from the regular morphology map. The outputs of the two branches are fused, and an initial image category label is output through a softmax classifier. The label is divided into two categories: benign (coded as 0) and osteosarcoma (preliminary identification, coded as 1). The preliminary classification result serves as the basic data structure for subsequent metabolic mapping correction steps. The label information and the map image are stored together in a database index for model updates and feature evaluation analysis.

[0032] Preferably, step S4, which involves analyzing the metabolite signal characteristics around the tumor based on benign tumor images and inferring the normal metabolic state of the tumor, includes: Multi-channel pixel mapping is performed on images of benign tumors to construct a multispectral pixel matrix of benign tumors; Based on the multispectral pixel matrix of benign tumors, neighborhood expansion and reconstruction are performed on the benign tumor image to generate an expanded image around the tumor. Extract metabolite signal features from the peritumor periphery of the extended image; Based on the signal characteristics of metabolites around the tumor, the salient region of tumor metabolites is located; Metabolic homeostasis patterns are mapped by saliency domains of tumor metabolites and metabolite signal characteristics, and the normal metabolic state of the tumor is determined.

[0033] In this embodiment, metabolic state analysis is performed on the acquired benign tumor image. First, multi-channel pixel mapping processing is performed on the benign tumor image. The original image is a three-channel RGB format image with a resolution of 256×256 pixels. This image is input to the spectral mapping module. This module stretches the three RGB channels into a 7-dimensional approximate narrowband spectral response based on a channel stretching transformation strategy. The R channel is mapped to the 630nm, 650nm, and 680nm bands, the G channel is mapped to the 510nm, 530nm, and 560nm bands, and the B channel is mapped to the 450nm band. The mapping method uses channel linear interpolation based on empirical coefficients and gamma adjustment. The gamma value is set to 1.2, and the linear interpolation factor is adjusted from 0.25 to 0.85 according to the channel weights. Finally, a benign tumor multispectral pixel matrix with a dimension of 256×256×7 is constructed. The multispectral pixel matrix is ​​input into the neighborhood expansion module, and the tumor boundary mask is extracted first. Information and boundary masks are output by the image segmentation network, using a U-Net network structure. After boundary extraction, the boundary is expanded by 12 pixels in each direction to form a neighborhood window. The pixel filling method within the neighborhood uses a mirror expansion method, symmetrically filling the image boundary pixels into the expansion area. After expansion, a channel preservation mechanism is used to simultaneously expand 7 channels, forming an expanded image of the tumor periphery with a size of 280×280×7. The expanded image of the tumor periphery is input into the metabolite signal extraction module. This module performs spectral signal deconstruction on the image based on a preset metabolite signal reference spectral library. The reference spectral library contains spectral response templates for three typical metabolites: choline (Cho), creatine (Cr), and N-acetylaspartate (NAA). Each metabolite has its typical reflectance curve in 7 channels. Spectral correlation matching is performed on the image at the pixel level. The cosine similarity between the spectral vector of each pixel and the reference spectrum is calculated. If the cosine similarity is greater than 0, the image is considered a valid metabolite.85 pixels were identified as response pixels of a certain metabolite. The response regions and intensities of the three metabolites in the entire extended image were statistically analyzed to form a metabolite signal feature vector around the tumor. Based on the metabolite signal feature vector, a metabolite response heatmap was generated in the image space. Independent response maps were generated for each of the three types of metabolites, and normalized to [0, 1] using the maximum response value. Then, a morphological dilation operation was performed on each response map, with a structuring element size of 3×3, for one iteration. Next, the area of ​​connected regions in each response map was calculated; regions with an area greater than 128 pixels were considered metabolic salient regions. The salient regions of the three types of metabolites were merged to form a tumor metabolite salient domain. Simultaneously, the centroid coordinates and boundary directions of each type of metabolite salient domain were statistically analyzed. Spatial overlap analysis was performed on the salient domains to determine the spatial consistency of metabolites. The metabolite salient domains were then compared with... Metabolic signal feature vectors are input to a metabolic state mapping module, which loads a three-dimensional metabolic homeostasis model. This model defines three metabolic ratio indices—Cho / Cr, NAA / Cr, and Cho / NAA—based on three-dimensional coordinate axes. The signal vectors are mapped to metabolic ratio vectors and then projected into points in three-dimensional space. If the projected point falls within a preset metabolic homeostasis region (volume boundaries are Cho / Cr < 1.3, NAA / Cr > 1.5, and Cho / NAA < 1.1), the current tumor metabolic state is determined to be normal. If the projected point exceeds any boundary, the output state is identified as non-steady-state, and the metabolic homeostasis determination label is output as "steady-state" or "non-steady-state." Simultaneously, the ratio data and determination label are combined into structured information for subsequent benign reference construction steps.

[0034] Preferably, step S4, combining the normal metabolic state of the tumor with images of benign tumors to form a benign reference group, includes: Temporal feature encoding is performed on the normal metabolic state data of the tumor to obtain the temporal code of the metabolic state; Structured image layering is performed on images of benign tumors to generate multi-layered image structures of benign tumors; The metabolic state temporal encoding and the multi-layer image structure of benign tumors are heterogeneously aligned to generate feature-aligned fusion data. The feature-aligned fusion data is enhanced by reference domain mapping, and a benign reference group is reorganized based on the enhanced reference domain.

[0035] In this embodiment, the process of constructing a benign reference group from the acquired normal metabolic state of the tumor and images of benign tumors first involves performing temporal feature encoding on the metabolic state data. The input data includes three ratio sequences: Cho / Cr, NAA / Cr, and Cho / NAA. All sequence data are sampled at 1-second intervals, with a sample length of 10 time segments. Within each time segment, a three-dimensional metabolic trajectory point sequence is constructed based on the collected ratio vectors. This point sequence is then input into a Bi-LSTM (Bidirectional Long Short-Term Memory) encoder for feature extraction. The network input dimension is 3, the hidden state dimension is 64, and the output is mapped through a fully connected layer to a temporal vector with a fixed dimension of 128. The output vector is named "Metabolic State Temporal". The encoding process encodes the trend fluctuations, stability changes, and signal response structure of metabolic ratios. Benign tumor images are input into a structured image layering module for hierarchical representation transformation. The original image dimensions are 256×256×3. A multi-resolution image sequence is constructed using an image pyramid generation mechanism, employing a Gaussian pyramid and Laplacian pyramid fusion method. Each pyramid image is convolved with a 5×5 Gaussian kernel and then downsampled by a factor of 2, resulting in a total of four image structures: 256×256, 128×128, 64×64, and 32×32. Further, a structural feature extractor is used to calculate the edge map, texture map, and color histogram of each layer, ultimately forming an image at each layer. The image structure, named the benign tumor multi-layer image structure, is composed of four layers of images and four layers of structural feature matrices. The metabolic state temporal encoding and the benign tumor multi-layer image structure are input into a heterogeneous feature alignment module. A cross-modal attention mechanism is used to achieve spatial and semantic alignment between the two modalities. This mechanism is based on a Transformer structure, using the metabolic state temporal encoding as the query vector and the flattened feature matrix of each layer in the image structure as the key-value vector. The encoder structure contains four multi-head attention layers, each with a dimension of 32. A corresponding alignment weight matrix is ​​generated for each layer's image features. A weighted fusion mechanism is used to reconstruct the image structure of each layer using channel weighting, and the outputs of all layers are concatenated. A unified fusion feature tensor with dimensions of 256×256×8 is formed, containing image structure channels and metabolic state modulation channels. This tensor serves as feature-aligned fusion data for the next step of mapping enhancement. The feature-aligned fusion data is input into the reference domain mapping enhancement module, which constructs an enhanced reference domain model. The model structure is a multi-scale context fusion network, containing two downsampling paths and one upsampling restoration path. These paths perform semantic compression, channel redistribution, and spatial response alignment operations on the fused image at different resolutions. In the downsampling path, a 3×3 convolution is used to extract spatial semantic information and reduce the number of channels to 4. In the upsampling path, a deconvolution layer is used to restore the image to a size of 256×256 and reconstruct the channels to 8 dimensions.Finally, a fused and enhanced benign image reference domain image is output. This image contains metabolic structure mapping, texture edge fusion, and signal consistency response, serving as the final enhanced reference domain data. The enhanced reference domain image is then combined with the original image input to a benign reference group construction module. This module jointly encodes the image and vectors. The image is input into a ResNet18 backbone network for image feature extraction, with an output dimension of 512. The metabolic state temporal encoding is mapped to a 512-dimensional vector using an MLP, and then element-wise added to the image feature vector and normalized to form the benign reference encoding. This encoding is stored in a database indexed by the benign sample number. The image structure, original metabolic ratio data, and regularity label fields are also retained as auxiliary fields, collectively forming the benign reference group data structure, which serves as the basic input source data structure for subsequent comparative analysis and atlas correction models.

[0036] Preferably, step S5 includes the following steps: Step S51: Map the feature point cloud of the benign reference group to obtain benign reference point cloud data; construct a three-dimensional tumor structure based on the tumor feature image; project the benign reference point cloud data onto the three-dimensional tumor structure to obtain projection plane data; Step S52: Based on the projection plane data, perform spatial anchoring and registration on the benign reference feature point cloud data and the three-dimensional tumor structure to obtain the corresponding spatial anchoring data; Step S53: Perform attribute vector transfer operation based on the spatial anchoring corresponding data to obtain the attribute vector transfer result; perform heterogeneous feature difference mapping on the attribute vector transfer result to obtain metabolic difference data; Step S54: Evaluate the metabolic state of the metabolic difference data according to the preset metabolic difference threshold to obtain the evaluated metabolic state; map the tumor feature image to the evaluated metabolic state to generate metabolic state mapping data. Step S55: Perform secondary category judgment on the tumor feature image based on metabolic state mapping data to generate metabolic mapping image category; perform category correction on the initial image category based on the metabolic mapping image category to identify osteosarcoma category.

[0037] In this embodiment, spatial coordinates and related attribute values ​​are extracted from the multidimensional feature encoding of the benign reference group. A PCL (Point Cloud Library)-based 3D point cloud generation algorithm is used to convert the encoded data into a 3D point cloud format. Each point in the point cloud contains XYZ coordinates and its corresponding metabolite signal intensity value. The number of sampling points is controlled at around 10,000 to ensure spatial detail. The point cloud coordinates are normalized to a unit cube space, completing the construction of the benign reference point cloud data. Simultaneously, a 3D tumor structure is generated based on the tumor feature image using deep learning. The structure construction employs a VoxelNet-based 3D voxelization method, estimating voxel depth information from the 2D image through a deep network and stacking them to generate a 512×512×128 3D voxel grid. The voxel values ​​in the grid reflect local tissue density. Subsequently, Marching is used... The Cubes algorithm extracts 3D isosurfaces to generate a 3D surface structure model of the tumor. Combined with point cloud data, benign reference point clouds are mapped to the surface corresponding to the 3D tumor structure using a projection matrix. A perspective projection algorithm is used to calculate the orthographic projection coordinates of the point clouds, obtaining projection plane data. The projection resolution matches the voxel resolution of the tumor structure. Based on the projection plane data, the Iterative Closest Point (ICP) is used... The Point Algorithm (ICP) performs spatial anchoring registration between benign reference feature point clouds and 3D tumor structure surface point clouds. During the ICP iteration process, a kd-tree is first used to accelerate the nearest neighbor search, with a maximum of 50 iterations and a convergence threshold of 1e-6 times the error reduction. During registration, the maximum rotation angle of the transformation matrix is ​​limited to ±15 degrees, and the scaling factor is limited to 0.95 to 1.05 to ensure stable and reasonable registration transformation. After registration, spatial anchoring data is output, containing the registration transformation matrix between the two point clouds and the corresponding point pair index information. Based on the spatial anchoring data, an attribute vector transfer operation is performed. This operation involves mapping metabolic attribute values ​​from the benign reference point cloud to the 3D tumor structure point cloud using the registration matrix. The transfer uses a bilinear interpolation method to weighted average the metabolic values ​​of neighboring points. The weights are determined by the inverse of the spatial distance. During the mapping process, points outside the range are filled with nearest neighbors to generate attribute vector transfer results. To enhance data consistency, the results are smoothed using multi-scale convolutional filters. The filter kernel sizes are 3×3×3, 5×5×5, and 7×7×7 layer by layer, with filter weights of 0.6, 0.3, and 0.1, respectively. After completion, heterogeneous feature difference mapping is performed on the results. The mapping algorithm is based on the gradient domain fusion method. The fusion input includes metabolic intensity difference, spatial gradient information, and local texture features. The mapping operation is performed by minimizing the total variation energy function. The final output is metabolic difference data with the same data dimension as the three-dimensional tumor structure, reflecting the distribution of local metabolic difference intensity. The metabolic state of the metabolic difference data is evaluated according to a preset metabolic difference threshold, which is set to 0 for metabolic difference intensity.25. Regions with voxel difference intensity exceeding the threshold are marked as metabolically abnormal areas. The overall voxel percentage of abnormal areas is calculated. If the abnormal percentage is less than 0.15, the state is considered metabolically stable; otherwise, it is considered metabolically abnormal. The metabolic state assessment results are stored as binary labels. Simultaneously, a metabolic state mapping map is generated based on the metabolic difference data. Metabolicly abnormal areas are highlighted in red and overlaid on the corresponding positions of the tumor feature images. The mapping process is achieved through backprojection from voxel coordinates to two-dimensional projected coordinates, ultimately generating a metabolic state mapping data image. Based on the metabolic state mapping data, a secondary classification judgment is performed on the tumor feature images. First, the metabolic state mapping data image is input into a classification network, which employs a fusion of convolutional neural networks and graph attention mechanisms. The network employs a composite structure with an input size of 256×256×4 (including RGB three channels and a metabolic mapping channel). It comprises five convolutional blocks and two graph attention layers. The graph attention layers weight the features of the metabolic mapping region. The output passes through a softmax layer to generate a multi-class probability distribution, classifying the image into two categories: benign and osteosarcoma. After the mapped image category label is output, a category correction operation is performed between it and the initial image category. The correction algorithm is based on a Bayesian inference model, combining the two probability distributions to calculate the final category label. If the two classification results match, the category is directly output; otherwise, the decision is made based on the principle of maximizing the posterior probability. After correction, the final osteosarcoma identification category label is output and stored in the classification result database.

[0038] Of particular importance, step S55 includes: Multi-scale convolutional analysis was performed on tumor feature images based on metabolic state mapping data, and metabolic response region features were extracted to obtain a metabolic response region feature map. Heterogeneous category attention fusion is performed based on the feature map of the metabolic response region, and the inter-class prior feature vectors corresponding to the initial image categories are combined to generate a category reconstruction probability distribution matrix. Based on the category-reconstructed probability distribution matrix, the tumor feature image is assigned the maximum probability category to obtain the metabolic mapping image category; The metabolic mapping image category and the initial image category are subjected to category consistency cross-discrimination processing, the category divergence vector is calculated and local minimum correction is performed to obtain the correction vector data; The initial image category is adjusted based on the correction vector data to generate the final identified osteosarcoma category.

[0039] In this embodiment, multi-scale convolutional analysis is performed on the metabolic state mapping data. First, the metabolic state mapping data image is input into a multi-scale convolutional neural network (CNN) module. This module contains three parallel convolutional paths, using convolutional kernels of sizes 3×3, 5×5, and 7×7 to extract spatial features at different scales. The convolutional stride is set to 1, and the padding strategy is "same" to maintain spatial dimensions. The number of output channels of each convolutional layer is set to 64. After convolution, batch normalization and ReLU activation are performed. Then, the feature maps output from the three convolutional paths are fused by pixel-wise weighted summation. The fusion weights are adaptively applied through a channel attention mechanism. The learning process generates a metabolic response region feature map with a size of 256×256×64, containing rich spatial distribution information of metabolic responses. This feature map is then input into a heterogeneous category attention fusion module. The module uses a multi-head self-attention mechanism to fuse heterogeneous category features. The inter-class prior feature vectors corresponding to the initial image categories are pre-obtained from the training set, with a dimension of 64. This prior feature vector serves as the query vector. The spatial location features of the feature map are used as keys and values ​​input into the multi-head attention mechanism. The number of attention heads is set to 8, and the hidden layer dimension is set to 512. The module output is the category reconstruction probability distribution. The matrix, with dimensions 256×256×2, corresponds to the probability distributions of benign and osteosarcoma categories. The probability values ​​for each pixel are summed to 1 for both categories. The probability matrix is ​​normalized using a softmax layer. Based on the reconstructed probability distribution matrix, a maximum probability category assignment operation is performed. For each pixel in the matrix, the category with the highest probability value is selected as the assigned category. Category labels are 0 for benign and 1 for osteosarcoma. The assignment process is performed at the pixel level. The result generates a metabolic mapping image category map, with the same size as the input feature map (256×256 pixels, single-channel). This category image serves as key input data for subsequent category consistency judgment. The mapped image category and the initial image category are cross-checked for category consistency. First, the image category labels of the two categories are converted into vector form. The category difference vector at corresponding positions is calculated. The difference vector is a binary vector, with 0 for the same category and 1 for different categories. Then, a local minimum correction algorithm is applied to the difference vector. The algorithm uses a sliding window of 7×7 pixels, traversing the entire image and calculating the sum of the difference values ​​within each window. If the sum is less than 3, all category difference labels within that window are corrected to 0, achieving smooth correction of local small-scale erroneous judgments. The corrected difference vector is used to construct correction vector data. The dimension of the correction vector data is the same as that of the input category image, and the value range is {0,}.1) Based on the correction vector data, the target category adjustment operation of the initial image category is performed. First, the proportion of pixels with a value of 1 in the correction vector is calculated. If the proportion exceeds 15%, the category adjustment mechanism is triggered, and the initial image category is globally flipped, changing the benign category to osteosarcoma. Otherwise, the original category remains unchanged. The category adjustment operation is completed by a discriminant model based on Bayesian probability inference. The model input is the initial category probability distribution and correction vector statistics, and the output is the adjusted final identification category. Finally, an osteosarcoma category identification label is generated. The label is stored as a global category value for the entire image, with an additional confidence score. The confidence score is calculated from the probability distribution of the corrected category and is used for subsequent classification result analysis and verification.

[0040] This invention also provides an intelligent recognition and classification system for osteosarcoma images based on an image recognition model, used to execute the intelligent recognition and classification method for osteosarcoma images based on an image recognition model as described above. The intelligent recognition and classification system for osteosarcoma images based on an image recognition model includes: The image preprocessing module is used to acquire tumor images; perform multi-sequence perturbation equalization enhancement on the tumor images to obtain enhanced block tumor images; and perform color space transformation and remapping on the enhanced block tumor images to generate color-normalized images. The feature reconstruction and analysis module is used to extract the tumor region image based on the color-normalized image, and to perform contrast inversion processing on the tumor region image to obtain the inverted tumor region image; and to reconstruct the tumor feature image based on the tumor region image and the inverted tumor region image. The edge rule classification module is used to identify the edge features of tumors in tumor feature images based on a preset image recognition model, and to determine the degree of regularity of the tumor edges based on the tumor edge features; the tumor feature images are initially classified and judged based on the degree of regularity of the tumor edges to generate an initial image category; The metabolic reference construction module is used to acquire images of benign tumors; analyze the signal characteristics of metabolites around the tumor based on the images of benign tumors, and infer the normal metabolic state of the tumor; and combine the normal metabolic state of the tumor and the images of benign tumors into a benign reference group. The metabolic correction and identification module is used to compare the metabolic state of tumor feature images with a benign reference group to obtain metabolic difference data; to map the metabolic state of tumor feature images based on the metabolic difference data to obtain metabolic state mapping data; and to perform category correction on the initial image category based on the metabolic state mapping data to identify the osteosarcoma category.

[0041] This invention utilizes an image preprocessing module to enhance acquired tumor images through multi-sequence perturbation equalization, resulting in enhanced, segmented tumor images with significantly improved detail and contrast. Color space transformation and remapping ensures image consistency under different imaging conditions, and the resulting color-standardized image effectively eliminates the influence of lighting differences between devices. A feature reconstruction and analysis module extracts tumor region images and performs contrast inversion processing, resulting in inverted tumor region images that effectively highlight key features. The reconstructed tumor feature image integrates the advantages of multiple processing methods, enhancing feature expressiveness. An edge rule classification module identifies edge features in the tumor feature image based on a preset image recognition model, significantly improving recognition accuracy. The determination of edge regularity provides a scientific basis for preliminary classification, reducing the influence of subjective factors. A metabolic reference construction module acquires benign tumor images and analyzes the surrounding metabolite signal characteristics, forming a benign reference group that provides a strong foundation for subsequent analysis. The inference of metabolic state can effectively capture normal... The metabolic feature and metabolic correction identification module uses a benign reference group to compare the metabolic state of tumor feature images. The generated metabolic difference data provides an important basis for further analysis. The metabolic state mapping based on the metabolic difference data not only improves the accuracy of image classification but also realizes dynamic correction of the initial category. The final identified osteosarcoma category has higher accuracy and reliability. The overall system achieves efficient processing of complex image data through the application of intelligent recognition technology, optimizes the image analysis process, improves data processing efficiency, reduces subjective interference in traditional methods, and enhances the objectivity and systematicness of data analysis. It provides new ideas for future research in related fields, promotes the in-depth application of intelligent recognition technology in the field of medical imaging, improves the ability to identify complex lesions such as osteosarcoma, promotes the advancement of imaging analysis technology, lays a solid foundation for the application of large-scale data processing scenarios, and promotes the combination of storage-computing technology and intelligent recognition technology, forming an efficient and accurate data processing system.

[0042] Therefore, the embodiments should be considered as exemplary and non-limiting in all respects, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of the equivalents of the application are intended to be included within the invention.

[0043] The above description is merely a specific embodiment of the present invention, enabling those skilled in the art to understand or implement the invention. Various modifications to these embodiments will be readily apparent to those skilled in the art, and the general principles defined herein may be implemented in other embodiments without departing from the spirit or scope of the invention. Therefore, the present invention is not to be limited to the embodiments shown herein, but is to be accorded the widest scope consistent with the principles and novel features of the invention herein.

Claims

1. A method for intelligent recognition and classification of osteosarcoma images based on an image recognition model, characterized in that, Includes the following steps: Step S1: Acquire images of the tumor; Multi-sequence perturbation equalization enhancement is performed on the tumor image to obtain enhanced block tumor images; Color space transformation and remapping are performed on the enhanced block tumor images to generate color-normalized images; Step S2: Extract the tumor region image from the color-normalized image, and perform contrast inversion processing on the tumor region image to obtain the inverted tumor region image; reconstruct the tumor feature image based on the tumor region image and the inverted tumor region image; Step S3: Identify the tumor edge features in the tumor feature image based on the preset image recognition model, and determine the regularity of the tumor edge based on the tumor edge features; perform preliminary classification judgment on the tumor feature image based on the regularity of the tumor edge to generate an initial image category; Step S4: Obtain images of benign tumors; analyze the signal characteristics of metabolites around the tumor based on the images of benign tumors, and infer the normal metabolic state of the tumor; combine the normal metabolic state of the tumor and the images of benign tumors into a benign reference group; Step S5: Compare the metabolic state of tumor feature images with the benign reference group to obtain metabolic difference data; map the metabolic state of tumor feature images based on the metabolic difference data to obtain metabolic state mapping data; correct the initial image category based on the metabolic state mapping data to identify the osteosarcoma category.

2. The intelligent recognition and classification method for osteosarcoma images based on an image recognition model according to claim 1, characterized in that, Step S1 includes the following steps: Step S11: Acquire tumor images; perform sequence splicing and fusion on the tumor images to obtain a multi-sequence tumor fusion image, wherein the number of spliced ​​sequences is limited to no less than 3 sets; inject perturbation information into the multi-sequence tumor fusion image data to generate a perturbation-enhanced tumor image; Step S12: The perturbation-enhanced tumor image is locally equalized and divided into blocks to generate a block equalized tumor image, wherein the size of each block is set to 32×32 pixels; the block equalized tumor image is subjected to multi-dimensional scale detail enhancement processing to obtain an enhanced block tumor image, wherein the enhancement scale level is set to 3-5 layers. Step S13: Perform edge-preserving smoothing processing on the enhanced block tumor image to generate an edge-enhanced smoothed tumor image; perform color space remapping on the edge-enhanced smoothed tumor image to obtain a color space-mapped tumor image; Step S14: Perform hue normalization processing on the tumor image based on color space mapping and perform pseudo-color encoding to obtain a pseudo-color tumor image, wherein the hue mapping range is set to 0°-270°. Step S15: Perform color layer resampling based on the pseudo-color tumor image to generate a color-normalized image.

3. The intelligent recognition and classification method for osteosarcoma images based on an image recognition model according to claim 1, characterized in that, Step S2, which involves extracting the tumor region image from the color-normalized image and performing contrast inversion processing on the tumor region image, includes: Identify tumor region boundaries based on color-standardized images; The tumor region boundary is expanded inward to obtain the expanded tumor region. The tumor region image is extracted from the color-normalized image by expanding the tumor region; The contrast inversion process is performed on the tumor region image to obtain the initial inverted tumor region image; Shape correction processing is performed on the inverted tumor region image to obtain the inverted tumor region image.

4. The intelligent recognition and classification method for osteosarcoma images based on an image recognition model according to claim 1, characterized in that, Step S2, which involves reconstructing the tumor feature image based on the tumor region image and the inverted tumor region image, includes: Affine registration is performed on the tumor region image and the reversed tumor region image, where the affine matrix is ​​limited to the rotation angle [-10°, +10°] and the scaling ratio [0.9, 1.1] to obtain the initial aligned image; Multi-channel residual enhancement is performed on the initial aligned image to generate a residual fused image; Guided edge filtering is applied to the residual fusion image to obtain a boundary optimized image, wherein the guide kernel size is set to 5×5 and the edge threshold range is set to [20, 50]. The boundary optimization image is subjected to morphological semantic compression mapping to obtain the tumor feature image.

5. The intelligent recognition and classification method for osteosarcoma images based on an image recognition model according to claim 1, characterized in that, Step S3 involves identifying tumor edge features in the tumor feature image based on a preset image recognition model, and determining the regularity of the tumor edge based on these features, including: Extract the tumor edge image from the tumor feature image based on a preset image recognition model; Remove false edges from the tumor edge image and extract the true contour of the tumor; Based on the true contour of the tumor, the edge features of the tumor are identified; The overall morphological analysis was performed based on the characteristics of the tumor margins, and the results were quantified into overall morphological characteristic indicators. Map the outer contour lines of the tumor to accurately depict its true outline; Based on the actual contour of the tumor, the outer contour lines are located by a sequence of equally spaced points to generate equally spaced contour points. Constructing a local neighborhood of contour points based on equally spaced contour points; Local undulation gradients of each subdomain are analyzed based on the local neighborhood of contour points. The regularity of the tumor margin is determined by assessing the overall morphological characteristics and the local undulation gradient of each sub-domain.

6. The intelligent recognition and classification method for osteosarcoma images based on an image recognition model according to claim 1, characterized in that, Step S3, which involves preliminary classification of the tumor feature image based on the regularity of the tumor edges, includes: The regularity data of tumor edges are expanded in feature dimensions to generate a multidimensional expression vector of regularity. The regularity-level multidimensional representation vector is subjected to feature clustering to obtain regular feature cluster labels; Identify rule structure anomalies in the rule feature cluster labels and mark them as rule structure anomaly labels; Graph structure encoding is performed on the regular feature cluster labels and structural anomaly labels to generate a comprehensive regular morphological encoding. The regular morphology is comprehensively encoded, then decoded as a feature, and mapped to a regular morphology map of the tumor. Based on the regular morphology atlas of tumors, a preliminary classification judgment is made on the tumor feature images to generate initial image categories.

7. The intelligent recognition and classification method for osteosarcoma images based on an image recognition model according to claim 1, characterized in that, Step S4, which involves analyzing the metabolite signal characteristics around the tumor based on benign tumor images and inferring the normal metabolic state of the tumor, includes: Multi-channel pixel mapping is performed on images of benign tumors to construct a multispectral pixel matrix of benign tumors; Based on the multispectral pixel matrix of benign tumors, neighborhood expansion and reconstruction are performed on the benign tumor image to generate an expanded image around the tumor. Extract metabolite signal features from the peritumor periphery of the extended image; Based on the signal characteristics of metabolites around the tumor, the salient region of tumor metabolites is located; Metabolic homeostasis patterns are mapped by saliency domains of tumor metabolites and metabolite signal characteristics, and the normal metabolic state of the tumor is determined.

8. The intelligent recognition and classification method for osteosarcoma images based on an image recognition model according to claim 1, characterized in that, Step S4, which combines images of the normal metabolic state of the tumor with images of benign tumors to form a benign reference group, includes: Temporal feature encoding is performed on the normal metabolic state data of the tumor to obtain the temporal code of the metabolic state; Structured image layering is performed on images of benign tumors to generate multi-layered image structures of benign tumors; The metabolic state temporal encoding and the multi-layer image structure of benign tumors are heterogeneously aligned to generate feature-aligned fusion data. The feature-aligned fusion data is enhanced by reference domain mapping, and a benign reference group is reorganized based on the enhanced reference domain.

9. The intelligent recognition and classification method for osteosarcoma images based on an image recognition model according to claim 1, characterized in that, Step S5 includes the following steps: Step S51: Map the feature point cloud of the benign reference group to obtain benign reference point cloud data; construct a three-dimensional tumor structure based on the tumor feature image; project the benign reference point cloud data onto the three-dimensional tumor structure to obtain projection plane data; Step S52: Based on the projection plane data, perform spatial anchoring and registration on the benign reference feature point cloud data and the three-dimensional tumor structure to obtain the corresponding spatial anchoring data; Step S53: Perform attribute vector transfer operation based on the spatial anchoring corresponding data to obtain the attribute vector transfer result; perform heterogeneous feature difference mapping on the attribute vector transfer result to obtain metabolic difference data; Step S54: Evaluate the metabolic state of the metabolic difference data according to the preset metabolic difference threshold to obtain the evaluated metabolic state; map the tumor feature image to the evaluated metabolic state to generate metabolic state mapping data. Step S55: Perform secondary category judgment on the tumor feature image based on metabolic state mapping data to generate metabolic mapping image category; perform category correction on the initial image category based on the metabolic mapping image category to identify osteosarcoma category.

10. An intelligent recognition and classification system for osteosarcoma images based on an image recognition model, characterized in that, For executing the intelligent recognition and classification method for osteosarcoma images based on an image recognition model as described in claim 1, the intelligent recognition and classification system for osteosarcoma images based on an image recognition model comprises: The image preprocessing module is used to acquire tumor images; perform multi-sequence perturbation equalization enhancement on the tumor images to obtain enhanced block tumor images; and perform color space transformation and remapping on the enhanced block tumor images to generate color-normalized images. The feature reconstruction and analysis module is used to extract the tumor region image based on the color-normalized image, and to perform contrast inversion processing on the tumor region image to obtain the inverted tumor region image; and to reconstruct the tumor feature image based on the tumor region image and the inverted tumor region image. The edge rule classification module is used to identify the edge features of tumors in tumor feature images based on a preset image recognition model, and to determine the degree of regularity of the tumor edges based on the tumor edge features; the tumor feature images are initially classified and judged based on the degree of regularity of the tumor edges to generate an initial image category; The metabolic reference construction module is used to acquire images of benign tumors; analyze the signal characteristics of metabolites around the tumor based on the images of benign tumors, and infer the normal metabolic state of the tumor; and combine the normal metabolic state of the tumor and the images of benign tumors into a benign reference group. The metabolic correction and identification module is used to compare the metabolic state of tumor feature images with a benign reference group to obtain metabolic difference data; to map the metabolic state of tumor feature images based on the metabolic difference data to obtain metabolic state mapping data; and to perform category correction on the initial image category based on the metabolic state mapping data to identify the osteosarcoma category.

Images