Method, apparatus, and storage medium for detecting discontinuous tissue content based on tomographic images

Abstract

This invention discloses a method, apparatus, and storage medium for detecting the content of discontinuous tissues based on tomographic images, belonging to the field of food quality testing technology. The invention provides a method for detecting the content of discontinuous tissues based on tomographic images, comprising: acquiring a tomographic image of the object to be tested; merging multiple non-target tissue pixels with overlapping gray levels into a single non-target category based on preset gray-level distribution features; processing the image using an image segmentation model that fuses global context and local detail features to generate a binary classification segmentation mask containing the target tissue and the single non-target category; and determining the target tissue mass percentage based on the mask voxel distribution and prior physical density. This application effectively eliminates multi-class boundary interference, improves the segmentation accuracy and content detection accuracy of discontinuous small target tissues, and achieves non-destructive quantitative detection.

Description

Technical Field

[0001] This invention relates to the field of food quality testing technology, and in particular to a method, apparatus and storage medium for detecting discontinuous tissue content based on tomographic images. Background Technology

[0002] Pork occupies a significant position in global meat consumption, accounting for more than one-third of total global meat consumption, highlighting its crucial role in the global dietary structure. With rising living standards and changing dietary habits, consumers are placing higher demands on pork quality. As a key nutritional indicator, fat content not only directly determines nutritional value but is also closely related to sensory characteristics such as tenderness, juiciness, and flavor, making it a core parameter for assessing meat quality and processing suitability. However, traditional methods for quantifying fat (Soxhlet extraction) are time-consuming, labor-intensive, and destructive, failing to meet the demands of the modern meat industry for rapid online analysis. Therefore, developing a non-destructive and rapid technology for determining the fat content in cut pork is crucial for improving the efficiency of quality assessment and promoting high-quality development in the meat industry.

[0003] A search revealed that Chinese patent application number 2020109992563 discloses a method for three-dimensional tissue segmentation and measurement based on deep neural networks. The method includes: acquiring CT images of live pigs, dividing them into training and testing sets, and labeling the training set; constructing a CT bed segmentation network and an internal organ segmentation network, and training them using the labeled training set to obtain CT bed segmentation models and internal organ segmentation models; predicting mask images using the CT bed segmentation model and the internal organ segmentation model, and removing the CT bed and internal organs; and extracting the fat, muscle, and bone components of the pig body from the CT images of the live pig, and calculating the overall mass of the pig body and the proportion of each tissue. This application provides a method for segmenting different continuous tissues of a pig, enabling the estimation of the tissue proportions of the whole pig. Using manual segmentation results as a reference, the accuracy of the proposed method's calculation results is evaluated, achieving acceptable quantitative precision on a macroscopic level. However, while the boundaries of different tissues in the CT images involved in the above application are continuous, accurate segmentation of discontinuous tissues in meat, especially the irregular distribution, small volume, and discontinuous nature of intramuscular fat in lean meat, presents challenges. In addition, for the fat content of cut meat, the conventional and acceptable method is to use Soxhlet extraction, which includes intramuscular fat.

[0004] The above content is only used to help understand the technical solution of the present invention and does not represent an admission that the above content is prior art. Summary of the Invention

[0005] The main objective of this invention is to provide a method, apparatus, and storage medium for detecting discontinuous tissue content based on tomographic images, aiming to solve the technical problem that traditional smartwatches in the prior art are repetitive and complicated to operate, and cannot achieve a large number of functions with simple operation commands.

[0006] To achieve the above objectives, the present invention provides a method for detecting discontinuous tissue content based on tomographic images, the method comprising the following steps: Acquire tomographic images of the object under test, wherein the object under test includes target tissue and non-target tissue; Based on preset grayscale distribution features, pixels of multiple non-target tissues with overlapping grayscale values ​​in the tomographic scan image are merged into a single non-target category; The tomographic scan image is processed by an image segmentation model to generate a binary segmentation mask that includes the category of the target tissue and the single non-target category; The image segmentation model is configured to fuse global contextual features and local detail features; Based on the category of the target tissue and the voxel distribution of the single non-target category in the binary classification segmentation mask, as well as the prior physical density of the target tissue and the non-target tissue, the mass proportion of the target tissue in the object under test is determined.

[0007] Optionally, the step of merging pixels of multiple non-target tissues with overlapping gray levels in the tomographic image into a single non-target category based on preset gray-level distribution features includes: Based on the gray-level cross-over region between the target tissue and the non-target tissue, a segmentation threshold range is determined; The pixels of the non-target tissue in the tomographic image that fall within the segmentation threshold range are merged with the pixels of other non-target tissues that do not fall within the segmentation threshold range to generate the single non-target category.

[0008] Optionally, determining the segmentation threshold range based on the grayscale overlap region between the target tissue and the non-target tissue includes: The median of the gray-level overlapping region is used as the boundary value of the segmentation threshold interval; The generation of a binary segmentation mask containing the category of the target organization and the single non-target category includes: The background pixels in the tomographic image are assigned a first label value, the pixels belonging to the category of the target tissue are assigned a second label value, and the pixels belonging to the single non-target category are assigned a third label value to generate the binary classification segmentation mask.

[0009] Optionally, the image segmentation model includes an encoder and a decoder; The encoder is configured to extract global contextual features of the tomographic scan image through a self-attention mechanism or dilated convolution; The decoder is configured to recover local detail features based on the global context features through a dense skip connection structure or a multi-scale feature fusion module to generate the binary classification segmentation mask.

[0010] Optionally, the encoder includes a Transformer module configured to capture long-range dependencies and combine convolutional operations to extract the global contextual features; or, The decoder includes a nested dense skip connection structure, which is configured to add multi-level feature fusion paths between the encoder and the decoder to narrow the semantic gap.

[0011] Optionally, before processing the tomographic image using an image segmentation model, the method further includes: The tomographic scan images are sequentially processed with speckle noise suppression, impulse noise removal, and global smoothing to suppress image artifacts while preserving tissue edge features.

[0012] Optionally, the step of sequentially performing speckle noise suppression processing, impulse noise removal processing, and global smoothing processing on the tomographic scan image includes: The tomographic images were processed using a Despeckle filter to suppress speckle noise; The image processed by the Despeckle filter is processed using a median filter to remove isolated impulse noise; A Gaussian filter is used to smooth the image after it has been processed by the median filter in order to suppress random noise.

[0013] Optionally, determining the mass proportion of the target tissue in the test object based on the category of the target tissue and the voxel distribution of the single non-target category in the binary classification segmentation mask, and the prior physical density of the target tissue and the non-target tissue includes: The number of first voxels representing the target tissue category and the number of second voxels representing the single non-target category are counted in the binary classification segmentation mask. Multiply the first number of voxels and the second number of voxels by the volume of a single voxel to obtain the first physical volume of the target tissue and the second physical volume of the non-target tissue. Based on the first physical volume, the second physical volume, the first prior physical density of the target tissue, and the second prior physical density of the non-target tissue, the mass percentage is calculated according to the following formula:

[0014] Where FMR is the mass percentage, V target For the first physical volume, ρ target V is the first prior physical density. non_target For the second physical volume, ρ non_target For the second prior physical density Furthermore, to achieve the above objectives, the present invention also proposes a device for detecting discontinuous tissue content based on tomographic images, the device comprising: An image acquisition module is used to acquire tomographic images of an object under test, wherein the object under test includes target tissue and non-target tissue. The category merging module is used to merge pixels of multiple non-target tissues with overlapping gray levels in the tomographic scan image into a single non-target category based on preset gray-level distribution features. An image segmentation module is used to process the tomographic scan image using an image segmentation model to generate a binary segmentation mask containing the category of the target tissue and the single non-target category; wherein, the image segmentation model is configured to fuse global context features and local detail features; The content determination module is used to determine the mass percentage of the target tissue in the test object based on the category of the target tissue and the voxel distribution of the single non-target category in the binary classification segmentation mask, as well as the prior physical density of the target tissue and the non-target tissue.

[0015] Furthermore, to achieve the above objectives, the present invention also proposes a storage medium storing a non-continuous tissue content detection program based on tomographic scan images. When the non-continuous tissue content detection program based on tomographic scan images is executed by a processor, it implements the steps of the non-continuous tissue content detection method based on tomographic scan images as described above.

[0016] This invention fundamentally simplifies the segmentation task by merging overlapping non-target tissues into a single category based on grayscale distribution features, eliminating the interference of multi-class boundary ambiguity on the identification of discontinuous small targets. Combined with an image segmentation model that integrates global context and local detail features, it ensures both the overall perception capability of diffusely distributed targets and maintains the boundary accuracy of minute structures. Furthermore, by jointly mapping voxel volume and prior physical density to calculate mass proportion, it overcomes the deficiency of traditional volume ratio methods that ignore tissue density differences. The synergistic effect of these three aspects significantly improves the accuracy and reliability of detecting discontinuous tissue content in tomographic images, providing an efficient, non-destructive, and high-precision solution for fields such as food quality assessment, medical image analysis, and industrial material testing. Attached Figure Description

[0017] Figure 1 This is a schematic diagram of the structure of a non-continuous tissue content detection device based on tomographic images, which is part of the hardware operating environment of the embodiment of the present invention. Figure 2 This is a schematic flowchart of the first embodiment of the method for detecting discontinuous tissue content based on tomographic images of the present invention; Figure 3 This is a structural block diagram of the first embodiment of the non-continuous tissue content detection device based on tomographic images of the present invention; Figure 4 The images are RGB images and X-ray CT images of two randomly selected pork samples for this invention. Figure 5 This invention describes the image preprocessing workflow and the processed CT image. Figure 6 This is a real label image of the present invention; Figure 7 This invention compares the predicted annotations generated by different deep learning methods with the manually annotated results. Figure 8 The images are the original images, manually annotated results, and difference diagrams of this invention. Figure 9 This invention uses FCBformer, UNet++, and DeeplabV3+ models to create scatter plots of the actual and predicted values ​​for the LMS(a–c) and FMR(d–f) datasets.

[0018] 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

[0019] It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0020] Reference Figure 1 , Figure 1 This is a schematic diagram of the structure of a non-continuous tissue content detection device based on tomographic images, which is part of the hardware operating environment of an embodiment of the present invention.

[0021] like Figure 1 As shown, the discontinuous tissue content detection device based on tomographic images may include: a processor 1001, such as a central processing unit (CPU), a communication bus 1002, a user interface 1003, a network interface 1004, and a memory 1005. The communication bus 1002 is used to enable communication between these components. The user interface 1003 may include a display screen or an input unit such as a keyboard; optionally, the user interface 1003 may also include a standard wired interface or a wireless interface. The network interface 1004 may optionally include a standard wired interface or a wireless interface (such as a Wireless-Fidelity (Wi-Fi) interface). The memory 1005 may be high-speed random access memory (RAM) or stable non-volatile memory (NVM), such as a disk drive. Optionally, the memory 1005 may also be a storage device independent of the aforementioned processor 1001.

[0022] Those skilled in the art will understand that Figure 1 The structure shown does not constitute a limitation on a device for detecting discontinuous tissue content based on tomographic images, and may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0023] like Figure 1 As shown, the memory 1005, which serves as a storage medium, may include an operating system, a network communication module, a user interface module, and a non-continuous tissue content detection program based on tomographic images.

[0024] exist Figure 1In the non-continuous tissue content detection device based on tomographic images shown, the network interface 1004 is mainly used for data communication with the network server; the user interface 1003 is mainly used for data interaction with the user; the processor 1001 and memory 1005 in the non-continuous tissue content detection device based on tomographic images of the present invention can be set in the non-continuous tissue content detection device based on tomographic images. The non-continuous tissue content detection device based on tomographic images calls the non-continuous tissue content detection program based on tomographic images stored in the memory 1005 through the processor 1001 and executes the non-continuous tissue content detection method based on tomographic images provided in the embodiment of the present invention.

[0025] This invention provides a method for detecting discontinuous tissue content based on tomographic images, referring to... Figure 2 , Figure 2 This is a schematic flowchart of the first embodiment of a method for detecting discontinuous tissue content based on tomographic images according to the present invention.

[0026] In this embodiment, the method for detecting discontinuous tissue content based on tomographic images includes the following steps: Step S10: Obtain a tomographic image of the object to be tested, wherein the object to be tested includes target tissue and non-target tissue.

[0027] It should be noted that the tomographic images can be three-dimensional volume data or two-dimensional slice sequences acquired by equipment such as X-ray computed tomography (CT), magnetic resonance imaging (MRI), or positron emission tomography (PET). The object to be analyzed refers to the physical sample requiring component analysis, such as chilled cut meat, live animal torsos, composite material components, or human organs. Target tissue refers to the specific component whose content needs to be identified and quantified in the current detection task, such as adipose tissue in pork, tiny lesions in medical images, or porosity defects in industrial parts; non-target tissue refers to all other components besides the target tissue, such as lean meat, skin, bones in pork, and background air within the scanning field of view. It should be understood that the definitions of target and non-target tissue are relative and depend on the specific detection purpose. The method in this embodiment is applicable to tomographic image analysis in any scene where there is grayscale overlap between discontinuous small targets and the surrounding medium.

[0028] It is worth noting that in this embodiment, pork was used as the cut meat sample, and a total of 34 cut meat samples from different parts were collected. The pig head was cut into cubic blocks with a length, width, and height of about 6 cm, and refrigerated at 4±0.5℃. The tests of each batch of samples, including X-ray CT scanning and fat content determination, were all completed on the same day. A desktop X-ray CT scanner was used to acquire CT image data of the pork samples. The equipment parameters were adjustable, with a rotating tube voltage of 50 kV and a current of 10 mA based on image contrast. The scanner was equipped with a 16-bit digital imaging acquisition system, and the generated image grayscale value (GSV) ranged from -32768 to 32767. The scanner used was the NAOMI-CT 002L from RF Corporation in Nagano Prefecture, Japan.

[0029] In the implementation, pork samples were placed at the center of rotation of the stage and scanned 360°. The software reconstructed and output 3D volumetric data, with 1100 consecutive XY tomographic slices for each sample, and a voxel resolution of 160×160×160 micrometers. A total of 35200 XY cross-sectional images with an effective pixel size of 900×900 were obtained from 34 samples. Among them, 22000 images from 20 randomly selected samples were divided into training and validation sets in a 7:3 ratio for training and optimizing the deep learning model. The remaining 15400 images from the 14 samples were used as an external prediction set to calculate the percentage of fat mass.

[0030] Step S20: Based on preset grayscale distribution features, merge pixels of multiple non-target tissues with overlapping grayscale values ​​in the tomographic scan image into a single non-target category.

[0031] It's important to note that in computed tomography (CT) imaging, due to volumetric effects, imaging noise, and the similarity of tissue physical properties, target tissue and surrounding non-target tissues often exhibit significant grayscale overlap. If traditional multi-classification strategies (e.g., simultaneously distinguishing between fat, lean meat, skin, and bone) are directly employed, the model needs to learn extremely subtle and fuzzy decision boundaries between these highly similar grayscale categories. This can easily lead to discontinuous small-volume targets (such as intramuscular fat) being misclassified as adjacent non-target tissues, or being missed due to boundary uncertainty. By forcibly merging multiple non-target tissues with overlapping grayscale into a single unified non-target category, the complex multi-classification problem is effectively simplified to a binary classification problem of "target tissue" versus "non-target background." This approach completely eliminates classification interference within non-target tissues, allowing the model to focus its computational power and feature representation capabilities on the differences between the target tissue and the overall non-target environment. This significantly improves the detection rate and segmentation integrity of discontinuous small targets in low-contrast regions. The preset grayscale distribution features can be prior knowledge obtained from historical data statistics, or dynamic parameters determined by online analysis of the current image histogram. This embodiment does not impose any restrictions on this.

[0032] Understandably, pork tissues of different densities exhibit varying degrees of X-ray attenuation, a characteristic that allows for contrast in grayscale-reconstructed cross-sectional images. Image contrast arises from the differences in X-ray absorption coefficients among different components within the sample, thus reflecting the spatial distribution differences of each tissue. High-density tissues attenuate X-rays more strongly, resulting in images with higher grayscale values; for example, medium-to-high grayscale areas could be categorized as muscle or skin tissue, while low grayscale areas represent adipose tissue. These grayscale differences among different pork tissues provide the technical capability for subsequent image segmentation and the calculation of voxels for different tissues.

[0033] In the specific implementation, to more intuitively demonstrate the density differences between different tissues, the experiment performed grayscale value statistical analysis on the acquired CT images. For each sample, three different 10×10×10 voxel samples of adipose and muscle tissue were selected, resulting in 30 voxels for each tissue type. Grayscale values ​​were extracted for each tissue, and segmentation thresholds for different tissues were calculated. The grayscale value of each voxel was also calculated. The results showed that the grayscale value of muscle tissue was significantly higher than that of adipose tissue, typically ranging from 400 to 1000, while the grayscale value of adipose tissue was concentrated between 0 and 400, consistent with their physical density characteristics. The grayscale value distribution of muscle tissue overlapped somewhat with that of adipose tissue due to the presence of some intramuscular fat in muscle, while the selected adipose tissue also contained a small amount of muscle tissue. Due to the specific composition and structure of the skin tissue, its low grayscale value distribution was 691±265, with most areas overlapping with the grayscale value of muscle tissue. Furthermore, the skeletons of different pork tissues did not exhibit the porous characteristics of fruits and vegetables; the tissue cells were dense and without gaps.

[0034] Step S30: Process the tomographic scan image using an image segmentation model to generate a binary segmentation mask containing the category of the target tissue and the single non-target category; wherein, the image segmentation model is configured to fuse global context features and local detail features.

[0035] It is worth noting that the image segmentation model can be any deep learning network with an encoder-decoder structure or equivalent feature extraction capabilities. Its core lies in simultaneously considering both global semantic understanding and fine-grained local reconstruction. For discontinuously distributed target tissues, while relying solely on local convolutional operations can capture texture details, it lacks an understanding of the overall spatial layout, easily misidentifying isolated noise points as targets. Conversely, relying solely on global downsampling features can understand macroscopic morphology but loses boundary information of minute targets. Therefore, this embodiment requires the model to fuse global contextual features and local detail features. The former is used to perceive the diffuse distribution pattern and long-distance dependencies of target tissues in the test object, while the latter is used to accurately recover the binary classification boundary simplified in step S200. The generated binary classification segmentation mask is a label matrix corresponding to the input image size, where each voxel or pixel is labeled as belonging to either the target tissue category or a single non-target category, without distinguishing specific subclasses of non-target tissues. This mask structure directly serves subsequent quantitative calculations, avoiding gaps or overlap errors that may occur when stitching multi-class masks.

[0036] In specific implementations, FCBformer is an image segmentation model based on the Transformer architecture. Its core design lies in introducing multi-scale feature fusion and context enhancement mechanisms. By combining convolutional operations with the Transformer's self-attention mechanism, this model effectively captures global dependencies and local details in images, improving the robustness of feature representation while maintaining high accuracy. FCBformer demonstrates good generalization ability and segmentation accuracy in tasks such as medical image segmentation and agricultural visual inspection. UNet++ is a semantic segmentation network developed based on the classic U-Net. Its core improvement lies in introducing a nested dense skip connection structure. This structure narrows the semantic gap and improves the model's ability to perceive targets at different scales by adding multi-level feature fusion paths between the encoder and decoder. UNet++ exhibits higher segmentation accuracy and stronger detail preservation ability in image segmentation tasks and is widely used in fields such as medical image analysis, industrial inspection, and agricultural visual recognition. DeepLabV3+ is a semantic segmentation framework based on dilated convolution. Its core design includes an encoder-decoder structure and a dilated spatial pyramid pooling module. The encoder part uses dilated convolution to expand the receptive field while maintaining resolution and capturing multi-scale contextual information; the decoder part achieves fine-grained boundary segmentation by fusing high-level semantic features and low-level detail features. This model achieves a good balance between segmentation accuracy and computational efficiency and is suitable for pixel-level classification tasks in complex scenes.

[0037] To compare the performance and consistency of deep learning methods for pork tissue segmentation, FCBformer, UNet++, and DeepLabv3+ were run independently 10 times each using the optimal hyperparameter combination, and the confusion matrix was output. In this study, the loss function (Loss), intersection-over-union ratio (IoU), Dice, sensitivity (SEN), specificity (SPE), and accuracy (ACC) were selected to evaluate the classifier performance, and the standard deviation (SD) was used to assess the consistency of the classifier predictions. One-way ANOVA was performed using SPSS statistical software (IBM, New York, USA), where p < 0.05 indicated statistical significance.

[0038] The three tables below compare the FMR predictions of the FCBformer, UNet++, and DeepLabv3+ models with the measured values ​​from the Soxhlet oil extraction method.

[0039]

[0040] Table 1 compares the FMR predicted values ​​from the FCBformer model with the measured values ​​from the Soxhlet oil extraction method.

[0041]

[0042] Table 2 compares the FMR predictions from the FUNet++ model with the measured values ​​from the Soxhlet oil extraction method.

[0043]

[0044] Table 3 compares the FMR predictions from the DeepLabv3+ model with the measured values ​​from the Soxhlet oil extraction method.

[0045] The Soxhlet extraction method was used to determine the fat content in pork samples and to calculate the total fat content of pork. The specific procedures were as follows: The pork sample was homogenized to form a mince, then dried in a forced-air drying oven at 105°C for 6 hours to completely remove moisture. The dried sample was then transferred to a Soxhlet extractor and extracted for at least 6 hours to fully extract the fat components. During the experiment, the initial mass of the mince, the mass after drying, the mass of the residue after extraction, and the mass of the extracted oil were recorded for subsequent calculation and analysis of the fat content.

[0046] To investigate the impact of image preprocessing methods on the performance of deep learning models, three deep learning models were trained using both the original images and images preprocessed with the aforementioned filters. The loss function value and Dice coefficient on the validation set were used as evaluation metrics for comparison. After preprocessing, the loss value of FCBFormer slightly increased from 0.141 to 0.180, indicating a slight increase in model convergence difficulty; the muscle Dice coefficient slightly decreased from 0.929 to 0.921, and the fat Dice coefficient decreased from 0.887 to 0.878, resulting in a slight negative impact on overall segmentation accuracy. After preprocessing, the loss value of UNet++ slightly decreased from 0.061 to 0.057, with a small fluctuation; the muscle Dice coefficient increased from 0.934 to 0.939, and the fat Dice coefficient increased from 0.883 to 0.899, showing a slight improvement in segmentation performance for both muscle and fat regions. DeepLabV3+ exhibited significant performance improvements after preprocessing, with the loss value decreasing from 0.230 to 0.169. The muscle Dice coefficient increased from 0.897 to 0.921, and the fat Dice coefficient increased significantly from 0.798 to 0.876, showing the most significant improvement. This indicates that the preprocessing strategy has a significant positive effect on the segmentation performance of DeepLabV3+, especially the segmentation accuracy of the fat region.

[0047] Step S40: Based on the category of the target tissue and the voxel distribution of the single non-target category in the binary classification segmentation mask, as well as the prior physical density of the target tissue and the non-target tissue, determine the mass proportion of the target tissue in the object to be tested.

[0048] It should be noted that existing technologies often use voxel count ratios or volume ratios to approximate tissue content, but this ignores the essential differences in physical density between different tissues. For example, in pork detection, the density of adipose tissue is significantly lower than that of lean meat tissue; calculating based solely on volume would lead to a systematic overestimation of the fat mass percentage. This embodiment establishes a precise mapping relationship from geometric spatial distribution to physical mass attributes by introducing a prior physical density parameter. The prior physical density refers to a known constant obtained based on a large number of sample statistics or literature records, rather than a real-time measurement of the current sample, which ensures the non-destructive and rapid nature of the method. In specific calculations, the total number of voxels of the target tissue category and the total number of voxels of a single non-target category in the binary classification segmentation mask are first counted. Combined with the voxel resolution of the tomographic image, this is converted into their respective physical volumes, and then multiplied by the corresponding prior physical density to obtain their respective mass estimates. Finally, the percentage of the target tissue in the total mass is calculated. This method fundamentally corrects the bias of pure geometric metrology, enabling non-destructive testing results to maintain a high degree of consistency with the gold standard of chemical analysis such as Soxhlet extraction, thus meeting the stringent requirements of industrial production and scientific research for quantitative accuracy.

[0049] It is worth noting that the Soxhlet extraction method was used to determine the fat content in pork samples, and the fat content of pork can be calculated. The specific operating steps are as follows: The pork sample was homogenized to make mince, and then placed in a forced-air drying oven at 105℃ for 6 hours to completely remove moisture; then the dried sample was transferred to a Soxhlet extractor, and extracted for at least 6 hours to fully extract the fat components. During the experiment, the initial mass of the mince, the mass after drying, the mass of the residue after extraction, and the mass of the extracted oil were recorded for subsequent calculation and analysis of fat content.

[0050] Reference Figures 4-6The process of merging pixels of multiple non-target tissues with overlapping gray levels in a tomographic image into a single non-target category based on preset gray-level distribution characteristics further includes: determining a segmentation threshold range based on the gray-level cross-over region between target and non-target tissues; and merging pixels of non-target tissues falling within the segmentation threshold range with pixels of other non-target tissues not falling within the segmentation threshold range to generate a single non-target category. Specifically, in tomographic imaging, due to the continuous transition of the attenuation coefficients of different biological tissues or material components to X-rays, coupled with the influence of partial volume effects, target tissues and their adjacent non-target tissues often do not exhibit a completely separated bimodal distribution in the gray-level histogram, but rather a significant gray-level cross-over region. Figure 5 In the diagram, the horizontal axis represents grayscale value, and the vertical axis represents relative frequency. The distribution curves of target and non-target tissues clearly overlap in the middle band. If a single fixed threshold is directly used for segmentation, pixels in the overlapping area are easily misclassified. This embodiment identifies this grayscale overlap area and determines a specific segmentation threshold range accordingly. Pixels of non-target tissues within this range, which were originally ambiguous in their classification, are forced into a single non-target category, thus eliminating the interference of grayscale ambiguity on subsequent model learning during the data preprocessing stage. This approach transforms the complex multi-class boundary partitioning problem into a clear interval mapping problem, significantly reducing the algorithm's sensitivity to slight grayscale differences and improving the robustness of the solution.

[0051] As a preferred implementation, when determining the segmentation threshold interval based on the gray-level overlap area between the target and non-target tissues, the median of the gray-level overlap area can be used as the boundary value of the segmentation threshold interval. Compared to using the mean, mode, or empirical fixed values, the median, as a statistical measure, has stronger resistance to interference from skewness and outliers in the gray-level distribution. In actual imaging, due to noise or local artifacts, the gray-level distribution in the overlap area may not be a perfect normal distribution. Using the median can more objectively characterize the probability equilibrium point of the two types of tissues in the overlap area, maximizing the balance between the two-way risks of misclassifying the target tissue as a non-target tissue and misclassifying the non-target tissue as the target tissue. It should be understood that although the median is preferred in this embodiment, in other implementations, a weighted average, a specific percentile, or a boundary value dynamically determined by a clustering algorithm can also be used, depending on the accuracy requirements of the specific application scenario, as long as it can characterize the central trend of the gray-level overlap area.

[0052] Furthermore, when generating a binary classification segmentation mask containing the target organization category and a single non-target category, the following labeling strategy is adopted: background pixels in the tomographic scan image are assigned a first label value, pixels belonging to the target organization category are assigned a second label value, and pixels belonging to the single non-target category are assigned a third label value to generate the binary classification segmentation mask. Specifically, this three-value label system (e.g., background = 0, single non-target category = 1, target organization = 2), although numerically represented as three levels, semantically serves the binary classification task. Assigning the same third label value to multiple non-target organizations with overlapping gray levels essentially erases the category boundaries between these non-target organizations within the label space. For deep learning models, this means that the loss function no longer penalizes confusion within non-target organizations, forcing the model to focus all its learning capabilities on distinguishing between the core decision boundary of "target organization" and "non-target whole". This label design effectively avoids the gradient oscillation problem caused by the blurred boundaries between non-target subclasses in traditional multi-class labeling, significantly improving the convergence speed and stability of model training, and is especially beneficial for improving the segmentation integrity of irregular, small-sized discontinuous target organizations.

[0053] To more clearly illustrate the above technical solution, the following explanation uses the detection of fat content in chilled, cut pork as an example. In this application scenario, the target tissue is fat, and non-target tissues include lean meat and pork skin. Figure 5 As shown, the gray values ​​of adipose tissue are mainly concentrated in the low gray value range, while the gray values ​​of lean meat and pork skin are higher and highly overlapping. Statistical analysis shows that the median of the gray value overlap area between adipose tissue and lean meat / skin tissue is 478. Therefore, in this embodiment, pixels with gray values ​​less than or equal to 478 are identified as potential target tissues or edges of overlapping areas, while pixels with gray values ​​greater than 478 (lean meat and pork skin) are merged into a single non-target category. When constructing training labels, the air background is marked as 0, the merged lean meat and pork skin area is marked as 1, and the adipose area is marked as 2. It should be noted that the above value of 478 is only an exemplary description for specific scanning parameters and sample types and does not constitute a limitation on the scope of protection of this application. In practical applications, this threshold boundary should be adaptively adjusted or recalibrated according to factors such as the specific material of the object under test, the tube voltage and current parameters of the imaging device, and the reconstruction algorithm. Through the above processing procedure based on median threshold and three-value label assignment, unnecessary classification interference between lean meat and pork skin was successfully eliminated, enabling the image segmentation model to accurately focus on the extraction of discontinuous small targets such as intramuscular fat, laying a reliable data foundation for subsequent high-precision quality ratio calculation.

[0054] The image segmentation model comprises an encoder and a decoder. Specifically, the encoder is configured to extract global contextual features from tomographic images using a self-attention mechanism or dilated convolution, while the decoder is configured to recover local detail features based on the global contextual features using a dense skip connection structure or a multi-scale feature fusion module to generate a binary segmentation mask. This encoder-decoder architecture is designed to address the dual need for both macroscopic spatial localization and precise microscopic boundary reconstruction of discontinuous small target tissues in tomographic images. The encoder is responsible for progressive downsampling to obtain high-level semantic information and understand the distribution pattern of the target tissue in the overall sample; the decoder is responsible for progressive upsampling, mapping abstract semantic features back to the original resolution to reconstruct a refined segmentation mask. Crucially, this embodiment requires the model to possess both global perception and local reconstruction capabilities simultaneously, without neglecting either. This is the structural foundation for achieving accurate segmentation of discontinuous tissues with overlapping grayscale values ​​and irregular shapes.

[0055] As a specific implementation, the encoder includes a Transformer module configured to capture long-range dependencies and combine convolutional operations to extract global contextual features. In computed tomography (CT) imaging, discontinuous target tissues (such as intramuscular fat) are often diffusely distributed throughout the sample, lacking fixed anatomical priors for their spatial location. Traditional pure convolutional neural networks, limited by fixed local receptive fields, struggle to establish feature associations across large spatial distances, easily misidentifying isolated noise as targets or missing small targets far from the main subject area due to a lack of global context. Introducing the Transformer module, its self-attention mechanism allows direct computation of the correlation between any two voxels in the image, effectively capturing long-range dependencies. Simultaneously, the convolutional operations preserve local texture details, enabling the encoder to both "see" the overall distribution of target tissues and "clearly perceive" local grayscale variations, significantly improving the model's ability to represent features of discontinuous targets in complex backgrounds.

[0056] As an alternative implementation, the decoder incorporates a nested dense skip connection structure. This structure is configured to add multi-level feature fusion paths between the encoder and decoder to bridge the semantic gap. In standard encoder-decoder networks, skip connections typically concatenate feature maps at the same level. However, due to multiple downsampling and nonlinear transformations by the encoder, a significant semantic gap exists between its deep features and the decoder's shallow features. Direct fusion often leads to the loss or confusion of detailed information. The nested dense skip connection structure forces a progressive fusion from coarse-grained semantics to fine-grained texture by constructing multiple cross-level feature transport paths between the encoder and decoder. This design is particularly beneficial for recovering the binary classification boundaries after dimensionality reduction in the aforementioned embodiments. It can losslessly transfer the global semantic information extracted by the encoder regarding "where the non-target organization is" to the shallow network in the decoder responsible for outlining the "target organization edges," thereby minimizing the blurring of boundaries for small targets while ensuring segmentation integrity.

[0057] Reference Figures 7-9 To illustrate the advantages of the above architecture more intuitively, such as Figure 8 As shown, the segmentation results of the model using the scheme in this embodiment and a pure convolutional model (such as DeepLabV3+) on the same pork CT slice are compared. It can be seen that the pure convolutional model, lacking effective global context modeling and multi-level detail recovery mechanisms, exhibits significant missed detections when handling discontinuous small targets such as intramuscular fat, incorrectly classifying some small fat regions as lean meat (i.e., a single non-target category), and its boundary prediction at tissue junctions is relatively coarse. In contrast, the model using a Transformer hybrid encoder or a nested dense skip connection decoder can not only accurately identify diffusely distributed small fat particles but also clearly delineate the complex boundaries between fat and lean meat. The generated binary classification segmentation mask highly matches the manually labeled ground truth. This comparison fully demonstrates that fusing global context features and local detail features is not a simple functional superposition but a necessary technical means to solve the problem of discontinuous tissue segmentation, and the resulting improvement in segmentation accuracy is unattainable by a pure convolutional architecture. It should be understood that although this embodiment lists two specific paths, Transformer and nested skip connections, in other implementations, any network variant that can achieve the same global-local feature fusion effect, such as a scheme that introduces a state space model or dynamic convolution mechanism, should be considered as an equivalent alternative within the scope of protection of this application, as long as it solves the same technical problem and achieves the expected technical effect.

[0058] Before processing the tomographic images using the image segmentation model, a preprocessing step is included. Specifically, speckle noise suppression, impulse noise removal, and global smoothing are performed sequentially on the tomographic images to suppress image artifacts while preserving tissue edge features. This preprocessing flow is not a simple filtering superposition, but a targeted cascade strategy based on the physical characteristics of tomographic imaging and the noise distribution patterns. Because discontinuous target tissues (such as intramuscular fat) are small in size and have blurred boundaries, any excessive smoothing or inappropriate denoising may cause the target signal to be submerged or morphologically distorted. Therefore, this embodiment constructs a progressive enhancement pipeline from specific denoising to general smoothing by strictly limiting the execution order of the three processes. This aims to provide clean and sharp input data for the subsequent image segmentation model, thereby indirectly improving the segmentation accuracy of small targets.

[0059] As a specific implementation, the steps executed sequentially above include: first, processing the tomographic image using a Despeckle filter to suppress speckle noise; then, processing the image after Despeckle processing using a median filter to remove isolated impulse noise; and finally, smoothing the image after median processing using a Gaussian filter to suppress random noise. Specifically, this particular execution order is irreversible, its underlying logic stemming from the hierarchical differences in noise types during image formation and the selectivity of each filter to the signal frequency response. The first step uses a Despeckle filter because speckle noise in tomographic images typically exhibits multiplicative noise and is highly coupled with tissue texture. The Despeckle algorithm can suppress speckles while preserving the edges and fine structures of the tissue to the greatest extent possible. If this step is performed later, the preceding smoothing operations will destroy the statistical characteristics of the speckles, leading to speckle removal failure or excessive edge blurring. The second step uses a median filter to remove isolated impulse noise (i.e., salt-and-pepper noise) introduced by detector dead pixels or reconstruction algorithms. The non-linear characteristics of the median filter allow it to remove extreme noise while protecting edges from being blunted. However, if performed before Despeckle, dense speckle noise may be misinterpreted as impulse signals, leading to loss of image details. The third step uses a Gaussian filter for global smoothing, mainly to suppress residual additive random noise and make the image grayscale distribution more continuous. Placing it last is to avoid the low-pass characteristic of the Gaussian kernel in the frequency domain prematurely smoothing out the weak boundaries of discontinuous small targets, ensuring that the edge information preserved in the first two steps is completely transmitted to the segmentation model.

[0060] To verify the effectiveness of this cascaded preprocessing flow, such as Figure 6The processing flowchart and its corresponding performance comparison results provide direct evidence, demonstrating the performance changes of three different image segmentation models before and after applying the preprocessing described in this embodiment. Experimental data show that this specific combination of preprocessing strategies has a significant effect on improving the segmentation accuracy of discontinuous tissues such as fat. For example, for the DeepLabV3+ model, after preprocessing, the Dice coefficient of its fat tissue increased significantly from 0.798 to 0.876, and the loss function value also decreased from 0.230 to 0.169, indicating that this preprocessing can effectively compensate for the lack of noise resistance in pure convolutional models. For the UNet++ model, which has superior performance, preprocessing also brought positive gains, with its fat Dice coefficient increasing from 0.883 to 0.899, and the muscle Dice coefficient also improving. Even the FCBformer model, which is more sensitive to preprocessing, maintained a high level of segmentation stability despite small fluctuations in various indicators. In contrast, using only a single filter or scrambling the above execution order cannot achieve the same effect, and may even lead to an increase in the false negative rate of small targets due to excessive edge smoothing. This fully demonstrates that the specific cascaded sequence of "Despeckle → Median → Gaussian" is the optimal solution that balances noise reduction and edge preservation, rather than a conventional method that can be arbitrarily replaced or adjusted by those skilled in the art. It should be understood that although this embodiment uses Despeckle, median, and Gaussian filters as examples for detailed explanation, in other embodiments, as long as the functional cascaded logic of "first speckle suppression, then impulse removal, and finally global smoothing" is followed, other variations of filtering algorithms with equivalent functions should also be considered equivalent technical solutions within the scope of this application.

[0061] In this embodiment, the process of determining the mass percentage of the target tissue in the test object based on the voxel distribution of the target tissue category and a single non-target category in the binary classification segmentation mask, as well as the prior physical density of the target tissue and non-target tissue, is further specified. Specifically, the process first counts the number of first voxels of the target tissue category and the number of second voxels of a single non-target category in the binary classification segmentation mask; then, the number of first voxels and the number of second voxels are multiplied by the volume of a single voxel, respectively, to obtain the first physical volume of the target tissue and the second physical volume of the non-target tissue; finally, based on the first physical volume, the second physical volume, the first prior physical density of the target tissue, and the second prior physical density of the non-target tissue, the mass percentage is calculated according to the following formula: .

[0062] Where FMR is the mass percentage, V target For the first physical volume, ρ target V is the first prior physical density. non_target For the second physical volume, ρ non_targetThis is the second prior physical density.

[0063] The core design of this calculation formula lies in the introduction of a "prior physical density" parameter, thereby establishing a precise mapping relationship from geometric spatial distribution to physical quality properties. Specifically, prior physical density refers to a known constant obtained based on a large number of sample statistics, literature records, or industry standards, rather than a value obtained through real-time destructive measurement of the current test object. For example, in the scenario of pork fat content detection, the prior physical density of adipose tissue is typically around 0.94 g / cm³, while the prior physical density of the non-target tissue, after combining lean meat and pork skin, is typically around 1.038 g / cm³. It should be understood that the above values ​​are only illustrative examples; in practical applications, adaptive calibration should be performed according to the specific type, variety, and imaging conditions of the test object, as long as it can characterize the average physical properties of the corresponding tissue. Since tomographic images themselves can only provide information on the geometric morphology of the tissue (i.e., the number of voxels), directly using the voxel number ratio or volume ratio to characterize the content actually implies the erroneous assumption that "all tissues have uniform density." However, in biological tissues or composite materials, the physical densities of different components often differ significantly. Taking pork as an example, the density of adipose tissue is significantly lower than that of lean tissue. If calculated solely by volume, the mass percentage of fat will be systematically overestimated, and vice versa, the content of high-density components will be underestimated. This systematic bias caused by differences in physical properties is an inherent defect that cannot be overcome by simply relying on image geometric features.

[0064] This embodiment utilizes the aforementioned formula to weight and correct volume data using prior physical density, fundamentally eliminating measurement errors caused by uneven tissue density. From a microscopic perspective, this formula essentially converts the "geometric fraction" obtained from image segmentation into a "mass fraction" that conforms to the law of conservation of mass. Experimental verification shows that the fat mass percentage (FMR) calculated using the formula in this embodiment has a high degree of consistency with the measured results of the gold standard chemical analysis such as Soxhlet extraction, with a coefficient of determination R² exceeding 0.92. In contrast, while the pure volume ratio method without density correction also shows some correlation with the true value, it exhibits a significant fixed deviation in absolute value, failing to meet the stringent quantitative accuracy requirements of industrial grading or precise nutritional assessment. Therefore, the method based on the joint calculation of voxel volume and prior density defined in this embodiment is not only a necessary technical means to achieve high-precision non-destructive testing but also a key substantive feature that distinguishes it from the coarse estimation methods in existing technologies. Furthermore, although this embodiment uses the calculation of fat mass percentage as an example for illustration, the calculation model is universal and can also be applied to the quantitative analysis of protein content, moisture content, or any other discontinuous components that can be distinguished by density and grayscale, simply by replacing the corresponding prior physical density parameters.

[0065] It should be understood that the above are merely illustrative examples and do not constitute any limitation on the technical solutions of the present invention. In specific applications, those skilled in the art can make settings as needed, and the present invention does not impose any restrictions on this.

[0066] It should be noted that the workflow described above is merely illustrative and does not limit the scope of protection of this invention. In practical applications, those skilled in the art can select some or all of the workflow to achieve the purpose of this embodiment according to actual needs, and no restrictions are imposed here.

[0067] Reference Figure 3 , Figure 3 This is a structural block diagram of the first embodiment of the non-continuous tissue content detection device based on tomographic images of the present invention.

[0068] like Figure 3 As shown, the discontinuous tissue content detection device based on tomographic images proposed in this embodiment of the invention includes: Image acquisition module 10 is used to acquire tomographic scan images of an object to be tested, the object to be tested including target tissue and non-target tissue; The category merging module 20 is used to merge pixels of multiple non-target tissues with overlapping gray levels in the tomographic scan image into a single non-target category based on preset gray-level distribution features. Image segmentation module 30 is used to process the tomographic scan image through an image segmentation model to generate a binary segmentation mask containing the category of the target tissue and the single non-target category; wherein, the image segmentation model is configured to fuse global context features and local detail features; The content determination module 40 is used to determine the mass proportion of the target tissue in the test object based on the category of the target tissue and the voxel distribution of the single non-target category in the binary classification segmentation mask, as well as the prior physical density of the target tissue and the non-target tissue. In addition, for technical details not described in detail in this embodiment, please refer to the method for detecting discontinuous tissue content based on tomographic images provided in any embodiment of the present invention, which will not be repeated here.

[0069] Furthermore, it should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0070] The sequence numbers of the above embodiments of the present invention are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0071] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of the present invention, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a storage medium (such as read-only memory (ROM) / RAM, magnetic disk, optical disk) and includes several instructions to cause a terminal device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of the present invention.

[0072] The above are merely preferred embodiments of the present invention and do not limit the scope of the patent. Any equivalent structural or procedural transformations made based on the description and drawings of the present invention, or direct or indirect applications in other related technical fields, are similarly included within the scope of patent protection of the present invention.

Claims

1. A method for detecting discontinuous tissue content based on tomographic images, characterized in that, include: Acquire tomographic images of the object under test, wherein the object under test includes target tissue and non-target tissue; Based on preset grayscale distribution features, pixels of multiple non-target tissues with overlapping grayscale values ​​in the tomographic scan image are merged into a single non-target category; The tomographic scan image is processed by an image segmentation model to generate a binary segmentation mask that includes the category of the target tissue and the single non-target category; wherein, the image segmentation model is configured to fuse global context features and local detail features; Based on the category of the target tissue and the voxel distribution of the single non-target category in the binary classification segmentation mask, as well as the prior physical density of the target tissue and the non-target tissue, the mass proportion of the target tissue in the object under test is determined.

2. The method for detecting discontinuous tissue content based on tomographic images as described in claim 1, characterized in that, The method of merging pixels of various non-target tissues with overlapping gray levels in the tomographic image into a single non-target category based on preset gray-level distribution features includes: Based on the gray-level cross-over region between the target tissue and the non-target tissue, a segmentation threshold range is determined; The pixels of the non-target tissue in the tomographic image that fall within the segmentation threshold range are merged with the pixels of other non-target tissues that do not fall within the segmentation threshold range to generate the single non-target category.

3. The method for detecting discontinuous tissue content based on tomographic images as described in claim 2, characterized in that, The step of determining the segmentation threshold range based on the grayscale overlap region between the target tissue and the non-target tissue includes: The median of the gray-level overlapping region is used as the boundary value of the segmentation threshold interval; The generation of a binary segmentation mask containing the category of the target organization and the single non-target category includes: The background pixels in the tomographic image are assigned a first label value, the pixels belonging to the category of the target tissue are assigned a second label value, and the pixels belonging to the single non-target category are assigned a third label value to generate the binary classification segmentation mask.

4. The method for detecting discontinuous tissue content based on tomographic images as described in claim 1, characterized in that, The image segmentation model includes an encoder and a decoder; The encoder is configured to extract global contextual features of the tomographic scan image through a self-attention mechanism or dilated convolution; The decoder is configured to recover local detail features based on the global context features through a dense skip connection structure or a multi-scale feature fusion module to generate the binary classification segmentation mask.

5. The method for detecting discontinuous tissue content based on tomographic images as described in claim 4, characterized in that, The encoder includes a Transformer module configured to capture long-range dependencies and combine convolutional operations to extract the global context features; or, The decoder includes a nested dense skip connection structure, which is configured to add multi-level feature fusion paths between the encoder and the decoder to narrow the semantic gap.

6. The method for detecting discontinuous tissue content based on tomographic images as described in claim 1, characterized in that, Before processing the tomographic image using the image segmentation model, the method further includes: The tomographic scan images are sequentially processed with speckle noise suppression, impulse noise removal, and global smoothing to suppress image artifacts while preserving tissue edge features.

7. The method for detecting discontinuous tissue content based on tomographic images as described in claim 6, characterized in that, The sequential processing of speckle noise suppression, impulse noise removal, and global smoothing on the tomographic scan image includes: The tomographic images were processed using a Despeckle filter to suppress speckle noise; The image processed by the Despeckle filter is processed using a median filter to remove isolated impulse noise; A Gaussian filter is used to smooth the image after it has been processed by the median filter in order to suppress random noise.

8. The method for detecting discontinuous tissue content based on tomographic images as described in claim 1, characterized in that, The determination of the mass proportion of the target tissue in the test object based on the voxel distribution of the target tissue category and the single non-target category in the binary classification segmentation mask, and the prior physical density of the target tissue and the non-target tissue, includes: The number of first voxels representing the target tissue category and the number of second voxels representing the single non-target category are counted in the binary classification segmentation mask. Multiply the first number of voxels and the second number of voxels by the volume of a single voxel to obtain the first physical volume of the target tissue and the second physical volume of the non-target tissue. Based on the first physical volume, the second physical volume, the first prior physical density of the target tissue, and the second prior physical density of the non-target tissue, the mass percentage is calculated according to the following formula: Where FMR is the mass percentage, V target For the first physical volume, ρ target V is the first prior physical density. non_target For the second physical volume, ρ non_target This is the second prior physical density.

9. A device for detecting discontinuous tissue content based on tomographic images, characterized in that, include: An image acquisition module is used to acquire tomographic images of an object under test, wherein the object under test includes target tissue and non-target tissue. The category merging module is used to merge pixels of multiple non-target tissues with overlapping gray levels in the tomographic scan image into a single non-target category based on preset gray-level distribution features. An image segmentation module is used to process the tomographic scan image using an image segmentation model to generate a binary segmentation mask containing the category of the target tissue and the single non-target category; wherein, the image segmentation model is configured to fuse global context features and local detail features; The content determination module is used to determine the mass percentage of the target tissue in the test object based on the category of the target tissue and the voxel distribution of the single non-target category in the binary classification segmentation mask, as well as the prior physical density of the target tissue and the non-target tissue.

10. A storage medium, characterized in that, The storage medium stores a non-continuous tissue content detection program based on tomographic scan images, which, when executed by a processor, implements the non-continuous tissue content detection method based on tomographic scan images as described in any one of claims 1 to 8.

Images