A hyperspectral technology-based intraoperative tumor rapid classification system

Abstract

The application discloses a kind of intraoperative tumor rapid classification system based on hyperspectral technology, comprising: image pre-processing module, for the hyperspectral image of several frequency bands of acquired tumor lesion tissue is pre-processed;Tumor lesion area identification module is used to select the hyperspectral image corresponding to multiple preferred frequencies respectively, fusion obtains fusion spectrum image, tumor target detection is carried out based on the image, and tumor lesion area is identified;Tumor lesion type identification module is used to carry out multi-pixel sampling to the tumor lesion area identified, obtains the frequency characteristics of each preferred frequency of each sampling point, and respectively input benign and malignant tumor identification model, respectively output benign and malignant possibility matrix, to identify tumor lesion type.The present application discriminates tumor benign and malignant based on tumor lesion tissue hyperspectral image, realizes the rapid classification of intraoperative tumor lesion type, reduces the rapid identification time, improves the accuracy of identification result while improving the identification efficiency.

Description

Technical Field

[0001] This invention relates to the field of hyperspectral technology applications, and more particularly to an intraoperative rapid tumor classification system based on hyperspectral technology. Background Technology

[0002] As is widely known, cancer is one of the most common diseases, and surgical treatment is widely applicable, mature, and one of the mainstream treatment methods. Elective surgery requires determining the nature of the tumor lesion intraoperatively to decide whether to remove the patient's organ or to determine the extent of the surgery, thereby deciding on the next steps in the surgical plan. In this process, the main means of intraoperative tumor detection currently includes intraoperative rapid frozen section pathology, which is usually only used to differentiate between benign and malignant tumors.

[0003] In existing techniques, the main steps of intraoperative rapid frozen section pathology examination include tissue freezing, sectioning, fixation, staining, mounting, and microscopic diagnosis. This diagnosis is obtained through visual observation and the pathologist's experience. Typically, a frozen section diagnosis report is available within 30 minutes of the surgical specimen arriving at the pathology department. This process is time-consuming, and waiting prolongs the surgery, increasing the burden on the patient and consuming medical resources. Furthermore, the pathologist's subjective judgment of the nature of the tumor lesion based on their own experience and knowledge significantly influences the diagnostic results, leading to relatively poor accuracy. Summary of the Invention

[0004] To address the shortcomings of existing technologies, this invention proposes an intraoperative rapid tumor classification system based on hyperspectral technology. This system determines the benign or malignant nature of tumors based on hyperspectral images of tumor lesions, enabling rapid intraoperative classification of tumor lesion types. Applied to intraoperative diagnosis, this system reduces rapid identification time, improves identification efficiency, and enhances the accuracy of identification results.

[0005] To achieve the above objectives, the present invention provides the following technical solution:

[0006] This invention provides an intraoperative rapid tumor classification system based on hyperspectral technology.

[0007] A rapid intraoperative tumor classification system based on hyperspectral technology, comprising:

[0008] The image preprocessing module is used to preprocess the acquired hyperspectral images of tumor lesions in several frequency bands;

[0009] The tumor lesion region identification module is used to select multiple hyperspectral images corresponding to preferred frequencies, fuse them to obtain a fused spectral image, and perform tumor target detection based on the fused spectral image to identify the tumor lesion region.

[0010] The tumor lesion type identification module is used to sample the identified tumor lesion area using multiple pixels, obtain the frequency characteristics of each preferred frequency of each sampling point, and input the frequency characteristics of each sampling point into the benign tumor identification model and the malignant tumor identification model respectively, and output the benign probability matrix and the malignant probability matrix respectively; based on the benign probability matrix and the malignant probability matrix, the type of tumor lesion is identified and determined.

[0011] Further technical solutions include preprocessing the acquired hyperspectral images, including:

[0012] Remove invalid frequency bands from several hyperspectral images, and then halve the gray values ​​of pixels in each frequency band that have a gray value greater than a set value to obtain the initial preprocessed hyperspectral image;

[0013] By removing invalid frequency bands from several hyperspectral reflectance images, the difference between the corresponding pixels in the initial preprocessed hyperspectral image and the hyperspectral reflectance image is calculated to obtain the preprocessed hyperspectral image.

[0014] A further technical solution, wherein the determination of the preferred frequency includes:

[0015] Acquire hyperspectral images of tumor lesions in several frequency bands, and crop the hyperspectral images to obtain hyperspectral images of the lesion area and its surrounding normal area;

[0016] Multi-pixel sampling is performed on the hyperspectral image of the lesion area to extract one-dimensional spectral characteristic data of each sampling point;

[0017] Multi-pixel sampling is performed on the hyperspectral image of the normal region to extract one-dimensional spectral characteristic data of each sampling point;

[0018] One-dimensional spectral characteristic data from multiple sampling points in the lesion area and one-dimensional spectral characteristic data from multiple sampling points in the normal area are randomly combined in pairs and the difference is calculated. The absolute values ​​of multiple sets of differences at each frequency are summed to obtain the sum of the absolute values ​​of multiple sets of differences at each frequency. The frequency corresponding to the maximum sum of absolute values ​​is selected as the preferred frequency.

[0019] A further technical solution involves fusing the selected hyperspectral images to obtain a fused spectral image. Specifically, the grayscale images corresponding to the selected multiple preferred frequencies of the hyperspectral images are used as the R, G, and B channels of the fused image for fusion, thereby achieving image fusion and obtaining a fused spectral image.

[0020] A further technical solution, namely, tumor target detection based on fused spectral images to identify tumor lesion areas, specifically includes:

[0021] Based on the fused spectral image, the trained target detection model is used to detect tumor targets, identify tumor lesion areas, and mark the tumor lesion areas with target detection anchor boxes.

[0022] The target detection model uses the YOLOv3 target detection algorithm and is trained using fused spectral images of several manually annotated tumor lesion regions as a training sample set to obtain the trained target detection model.

[0023] A further technical solution involves sampling the identified tumor lesion area using multiple pixels to obtain the frequency characteristics of each preferred frequency at each sampling point, specifically as follows:

[0024] The tumor lesion area marked by the target detection anchor frame is determined. Taking the center of the anchor frame as the reference point, this reference point is taken as point 0, the midpoint from point 0 to the boundary of the anchor frame is taken as points 1-4, and the midpoint from points 1-4 to point 0 is taken as points 5-8. Multi-pixel sampling is performed on these 8 points 1-8 to obtain the frequency characteristics of each sampling point at each preferred frequency.

[0025] A further technical solution is that the benign tumor identification model uses a convolutional neural network to connect a fully connected layer and a binary classifier to distinguish benign tumors from others;

[0026] The training process of the benign tumor identification model is as follows: using the one-dimensional frequency characteristic data of multiple sampling points of the hyperspectral image of the benign tumor lesion area as the training sample dataset, the model is trained to obtain the identification model for benign tumors or other benign tumors.

[0027] A further technical solution is that the malignant tumor identification model uses a convolutional neural network and a binary classifier to distinguish between malignant tumors and others;

[0028] The training process of the malignant tumor identification model is as follows: using the one-dimensional frequency characteristic data of multiple sampling points of the hyperspectral image of the malignant tumor lesion area as the training sample dataset, training is performed to obtain an identification model for malignant tumors or other malignant tumors.

[0029] A further technical solution involves identifying and determining the type of tumor lesion based on a benign probability matrix and a malignancy probability matrix, specifically as follows:

[0030] Based on the benign and malignant probability matrices, a probability matrix analysis is performed:

[0031] If the benign probability matrix is ​​benign and the malignant probability matrix is ​​other, then the final tumor lesion type is determined to be a benign tumor.

[0032] If the benign probability matrix is ​​"other" and the malignant probability matrix is ​​"malignant", then the final tumor lesion type is determined to be a malignant tumor.

[0033] If the benign probability matrix is ​​"other" and the malignant probability matrix is ​​also "other", then the identified tumor lesion area is determined to be normal tissue or other non-tumor lesions.

[0034] If the benign probability matrix is ​​benign and the malignant probability matrix is ​​malignant, further processing is still required.

[0035] A further technical solution is to input the one-dimensional frequency characteristic data of the identified tumor lesion region into the trained classification model to determine the type of the tumor lesion.

[0036] The above technical solutions have the following beneficial effects:

[0037] 1. The present invention proposes an intraoperative rapid tumor classification system based on hyperspectral technology, which determines the benign or malignant nature of tumors based on hyperspectral images of tumor lesions, realizes rapid classification of tumor lesion types during surgery, and is applied to intraoperative diagnosis, reducing rapid identification time, improving identification efficiency, and improving the accuracy of identification results.

[0038] 2. The solution described in this invention does not require a complicated slide preparation process. It directly obtains hyperspectral images of the cut edge or surface of the excised tissue. All operations can be completed in the operating room, reducing the time for rapid intraoperative diagnosis of tumors from 30 minutes to 10 minutes, which greatly saves time and costs.

[0039] 3. The system described in this invention does not require human judgment or identification, and the final identification result is not affected by the doctor's personal factors, which improves the accuracy and objectivity of the identification result and saves manpower costs and medical resources. Attached Figure Description

[0040] The accompanying drawings, which form part of this invention, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an improper limitation of the invention.

[0041] Figure 1 This is a schematic diagram of the intraoperative tumor rapid classification system in Embodiment 1 of the present invention;

[0042] Figure 2 This is an overall flowchart of the intraoperative tumor rapid classification system in Embodiment 1 of the present invention;

[0043] Figure 3 This is a flowchart of the hyperspectral image preprocessing in Embodiment 1 of the present invention;

[0044] Figure 4 This is a flowchart of determining the preferred frequency band in Embodiment 1 of the present invention. Detailed Implementation

[0045] It should be noted that the following detailed descriptions are exemplary and intended to provide further illustration of the invention. Unless otherwise specified, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this invention pertains.

[0046] It should be noted that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention. As used herein, the singular form is intended to include the plural form as well, unless the context clearly indicates otherwise. Furthermore, it should be understood that when the terms "comprising" and / or "including" are used in this specification, they indicate the presence of features, steps, operations, devices, components, and / or combinations thereof.

[0047] This embodiment proposes an intraoperative rapid tumor classification system based on hyperspectral technology, such as... Figure 1 As shown, it includes an image preprocessing module, a tumor lesion region identification module, and a tumor lesion type identification module:

[0048] The image preprocessing module is used to preprocess the acquired hyperspectral images of tumor lesions in several frequency bands;

[0049] The tumor lesion region identification module is used to select multiple hyperspectral images corresponding to preferred frequencies, fuse them to obtain a fused spectral image, and perform tumor target detection based on the fused spectral image to identify the tumor lesion region.

[0050] The tumor lesion type identification module is used to sample the identified tumor lesion area using multiple pixels, obtain the frequency characteristics of each preferred frequency of each sampling point, and input the frequency characteristics of each sampling point into the benign tumor identification model and the malignant tumor identification model respectively, and output the benign probability matrix and the malignant probability matrix respectively; based on the benign probability matrix and the malignant probability matrix, the type of tumor lesion is identified and determined.

[0051] In this embodiment, as Figure 2 As shown, firstly, the hyperspectral images are preprocessed using an image preprocessing module. Before preprocessing, several hyperspectral images of the intraoperative tumor lesion tissue are acquired. Specifically, using remote sensing technology, an imaging spectrometer simultaneously images the slice in multiple continuous and subdivided spectral bands in the ultraviolet, visible, near-infrared, and mid-infrared regions of the electromagnetic spectrum, obtaining hyperspectral images in several frequency bands. Simultaneously with acquiring the hyperspectral images, hyperspectral reflectance images of the intraoperative tumor lesion tissue are also acquired. The hyperspectral images are directly captured using a hyperspectral camera, while the hyperspectral reflectance images are obtained by emitting a beam of light, allowing it to be reflected by the tissue, and then capturing the image with a hyperspectral camera.

[0052] Next, the acquired hyperspectral images are preprocessed, such as... Figure 3As shown, it includes:

[0053] The invalid frequency bands before 410nm in the hyperspectral image are removed. Then, the gray values ​​of pixels with a gray value greater than the gray value set to 10 in the hyperspectral image of each frequency band are halved to obtain the initial preprocessed hyperspectral image.

[0054] The invalid frequency band before 410nm in the hyperspectral reflectance image is removed. The difference between the corresponding pixels in the initial preprocessed hyperspectral image and the hyperspectral reflectance image is calculated to obtain the preprocessed hyperspectral image.

[0055] The above-mentioned ambient light correction preprocessing scheme weakens the spectral information of ambient light and enhances the spectral characteristics of reflected light through physical isolation and image processing, thereby eliminating the interference of ambient light on the spectral characteristics of hyperspectral images and ensuring the accuracy of subsequent image recognition.

[0056] After obtaining the preprocessed hyperspectral image, it is input into the tumor lesion region identification module for tumor lesion region identification. Specifically, multiple hyperspectral images corresponding to preferred frequencies are selected and fused to obtain a fused spectral image. Based on the fused spectral image, tumor target detection is performed to identify tumor lesion regions.

[0057] Before performing the above steps for the first time, it is necessary to determine the preferred frequencies. In this embodiment, three frequencies are identified where the spectral characteristics of the hyperspectral image of the lesion tissue are significantly different from those of the hyperspectral image of the normal tissue. These three frequencies are selected as the preferred frequencies. Specifically, as follows... Figure 4 As shown, it includes the following steps:

[0058] Hyperspectral images of several frequency bands of tumor lesion tissue sections are obtained, and the hyperspectral images are cropped to obtain hyperspectral images of the lesion area and its surrounding normal area.

[0059] Multiple pixel points are sampled from the hyperspectral image of the lesion area. In this embodiment, 24 points are sampled. If the image size does not meet the 24-point sampling requirement, the number of sampling points can be reduced, and one-dimensional spectral characteristic data of each sampling point can be extracted.

[0060] Similarly, multi-pixel sampling is performed on the hyperspectral image of the normal region. In this embodiment, 24 points are sampled. If the image size does not meet the 24-point sampling requirement, the number of sampling points can be reduced, and one-dimensional spectral characteristic data of each sampling point can be extracted.

[0061] One-dimensional spectral characteristic data from multiple sampling points in the lesion area and one-dimensional spectral characteristic data from multiple sampling points in the normal area are randomly paired and their differences are calculated. The absolute values ​​of the 24 sets of differences at each frequency are summed to obtain the sum of the absolute values ​​of the 24 sets of differences at each frequency. The frequencies corresponding to the top 3 sums of absolute values ​​show significant differences between the tumor lesion area and the normal area in the image, and these 3 frequencies are selected as the preferred frequencies. In this embodiment, the preferred frequencies are confirmed to be 640nm, 550nm, and 470nm.

[0062] After determining the preferred frequencies, the hyperspectral images corresponding to the three preferred frequencies are selected, and these three hyperspectral images are fused to obtain a fused spectral image. Here, fusion refers to using hyperspectral software or MATLAB programming to use the corresponding grayscale images of the three selected frequencies as the R, G, and B channels of the fused image for fusion, thereby achieving image fusion and obtaining the fused spectral image.

[0063] Based on the fused spectral image, a trained target detection model is used to detect tumor targets and identify tumor lesion regions. Simultaneously, for each identified tumor lesion region, considering that multiple anchor boxes will be drawn during target detection, each anchor box, while defining the detection area, also indicates the probability that the object within that area is the target object. This probability is calculated by the target detection algorithm. Tumor lesion regions with a probability greater than 0.85 marked by the target detection anchor boxes are selected and labeled individually. This target detection model uses the YOLOv3 target detection algorithm and is trained using the fused spectral image of manually annotated tumor lesion regions as training samples to obtain the trained target detection model.

[0064] In reality, tumor target detection using object detection models may also have errors, meaning the object detection model may fail to identify the tumor lesion area. In this case, the following steps are also taken: determining whether the object detection model has detected the tumor lesion area. If no tumor lesion area is detected, the object detection model outputs the tumor lesion area marked by the anchor box; if no tumor lesion area is detected, the object detection model outputs the original fused spectral image, and the tumor lesion area in the image is manually selected. The most suitable anchor box is selected manually and labeled one by one. The manual selection avoids the situation where the object detection model identifies the lesion incorrectly.

[0065] The tumor lesion area marked by the anchor box is input into the tumor lesion type identification module. The tumor lesion type identification module performs multi-pixel sampling on the identified tumor lesion area, obtains the frequency characteristics of each optimal frequency of each sampling point, and inputs the frequency characteristics of each sampling point into the benign tumor identification model and the malignant tumor identification model respectively, and outputs the benign probability matrix and the malignant probability matrix respectively. Based on the benign probability matrix and the malignant probability matrix, the type of tumor lesion is identified and determined, and finally the type of tumor lesion is output.

[0066] Specifically, the tumor lesion area marked by the target detection anchor frame is determined. Taking the center of the anchor frame as the reference point, this reference point is taken as point 0, the midpoint from point 0 to the boundary of the anchor frame is taken as points 1-4, and the midpoint from points 1-4 to point 0 is taken as points 5-8, for a total of 8 points. Multi-pixel sampling is performed to obtain the frequency characteristics of each sampling point at each preferred frequency, that is, to obtain the one-dimensional frequency characteristic data of the 8 sampling points. The one-dimensional frequency characteristic data includes the 3 frequency characteristic data of the sampling point.

[0067] The one-dimensional frequency characteristic data of each sampling point are input into the benign tumor identification model and the malignant tumor identification model, respectively, and the benign probability matrix and the malignant probability matrix are output respectively.

[0068] The aforementioned benign tumor identification model uses a convolutional neural network, DenseNet201. After DenseNet201, a fully connected layer and a Softmax binary classifier are connected in sequence to distinguish between benign tumors and others. The training process of this benign tumor identification model is as follows: the one-dimensional frequency characteristic data of multiple sampling points of the hyperspectral image of the benign tumor lesion area is used as the training sample dataset for training, and the model for identifying benign tumors or other benign tumors is obtained.

[0069] The aforementioned malignant tumor identification model uses a convolutional neural network DenseNet101, followed by a Softmax binary classifier to distinguish between malignant tumors and others. The training process of this malignant tumor identification model is as follows: the one-dimensional frequency characteristic data of multiple sampling points of the hyperspectral image of the malignant tumor lesion area is used as the training sample dataset for training, and the model for identifying malignant tumors or other malignant tumors is obtained.

[0070] The aforementioned probability matrix is ​​the output of the Softmax binary classifier. For example, in the benign tumor identification model, when using Softmax binary classification, the output probability matrix contains two elements: [probability of benign tumor, probability of other]. The probability with the highest probability is taken as the classification output result. Similarly, in the malignant tumor identification model, when using Softmax binary classification, the output probability matrix contains two elements: [probability of malignant tumor, probability of other]. The probability with the highest probability is taken as the classification output result.

[0071] In step 4, the final tumor lesion type is determined based on the benign probability matrix and the malignant probability matrix.

[0072] In this embodiment, probability matrix analysis is performed based on the benign probability matrix and the malignant probability matrix. This involves evaluating the benign and malignant classification results of the tumor lesion area. Let the benign probability matrix be c1 and the malignant probability matrix be c2. Using 0.85 as the standard, if the benign probability matrix c1 is benign and the malignant probability matrix c2 is other, then the final tumor lesion type is determined to be a benign tumor. If the benign probability matrix c1 is other and the malignant probability matrix c2 is malignant, then the final tumor lesion type is determined to be a malignant tumor. If the benign probability matrix c1 is other and the malignant probability matrix c2 is also other, then the identified tumor lesion area is determined to be normal tissue or other non-tumor lesions. If the benign probability matrix c1 is benign and the malignant probability matrix c2 is malignant, further processing is still required. That is, the one-dimensional frequency characteristic data of the identified tumor lesion area is input into the trained classification model to determine the type of the tumor lesion.

[0073] The aforementioned standard of 0.85 means that if the largest element in the probability matrix is ​​greater than or equal to 0.85, the matrix is ​​considered valid and retained for further judgment. In this embodiment, the 0.85 standard refers to the AUC area standard, which generally considers results greater than 0.85 to be more reliable. Typically, a good model outputs a probability matrix with a probability of approximately 0.99 for the detected object's classification. However, in this embodiment, due to the use of a mechanism utilizing two detection channels for judgment, 0.85 is chosen as the standard to ensure the smooth completion of the probability matrix analysis.

[0074] The above classification model uses a residual neural network ResNet152, followed by a Softmax tri-classifier to distinguish between benign tumors, malignant tumors, and others. The training process of this classification model is as follows: the one-dimensional frequency characteristic data of multiple sampling points of hyperspectral images of benign and malignant tumor lesions are used as training sample datasets to train the model and obtain a classification model that can identify benign tumors, malignant tumors, or others.

[0075] The aforementioned intraoperative tumor rapid classification system based on hyperspectral technology distinguishes between benign and malignant tumors based on hyperspectral images of tumor lesions, enabling rapid intraoperative classification of tumor lesion types. This avoids the negative impact of manual identification of tumor lesion types. When applied to intraoperative diagnosis, it reduces rapid diagnosis time, improves diagnostic efficiency, and enhances the accuracy of diagnostic results.

[0076] The above description is merely a preferred embodiment of the present invention and is not intended to limit the invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

[0077] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

[0078] While the specific embodiments of the present invention have been described above in conjunction with the accompanying drawings, this is not intended to limit the scope of protection of the present invention. Those skilled in the art should understand that various modifications or variations that can be made by those skilled in the art without creative effort based on the technical solutions of the present invention are still within the scope of protection of the present invention.

Claims

1. A rapid intraoperative tumor classification system based on hyperspectral technology, characterized in that, include: The image preprocessing module is used to preprocess the acquired hyperspectral images of tumor lesions in several frequency bands; The tumor lesion region identification module is used to select hyperspectral images corresponding to three preferred frequencies, fuse them to obtain a fused spectral image, and perform tumor target detection based on the fused spectral image to identify the tumor lesion region. The determination of the preferred frequency includes: Acquire hyperspectral images of tumor lesions in several frequency bands, and crop the hyperspectral images to obtain hyperspectral images of the lesion area and its surrounding normal area; Multi-pixel sampling is performed on the hyperspectral image of the lesion area to extract one-dimensional spectral characteristic data of each sampling point; Multi-pixel sampling is performed on the hyperspectral image of the normal region to extract one-dimensional spectral characteristic data of each sampling point; One-dimensional spectral characteristic data from multiple sampling points in the lesion area and one-dimensional spectral characteristic data from multiple sampling points in the normal area are randomly combined in pairs and the difference is calculated. The absolute values ​​of multiple sets of differences at each frequency are summed to obtain the sum of the absolute values ​​of multiple sets of differences at each frequency. The frequencies corresponding to the sums of the top 3 absolute values ​​are selected as the preferred frequencies. The tumor lesion type identification module is used to sample the identified tumor lesion area using multiple pixels, obtain the frequency characteristics of each preferred frequency of each sampling point, and input the frequency characteristics of each sampling point into the benign tumor identification model and the malignant tumor identification model respectively, and output the benign probability matrix and the malignant probability matrix respectively; based on the benign probability matrix and the malignant probability matrix, the type of tumor lesion is identified and determined.

2. The intraoperative rapid tumor classification system based on hyperspectral technology as described in claim 1, characterized in that, The acquired hyperspectral images are preprocessed, including: Remove invalid frequency bands from several hyperspectral images, and then halve the gray values ​​of pixels in each frequency band that have a gray value greater than a set value to obtain the initial preprocessed hyperspectral image; By removing invalid frequency bands from several hyperspectral reflectance images, the difference between the corresponding pixels in the initial preprocessed hyperspectral image and the hyperspectral reflectance image is calculated to obtain the preprocessed hyperspectral image.

3. The intraoperative rapid tumor classification system based on hyperspectral technology as described in claim 1, characterized in that, The fusion process yields a fused spectral image. Specifically, the grayscale images corresponding to the three selected preferred frequencies of the hyperspectral images are used as the R, G, and B channels of the fused image for fusion, thereby achieving image fusion and obtaining the fused spectral image.

4. The intraoperative rapid tumor classification system based on hyperspectral technology as described in claim 1, characterized in that, The tumor target detection based on fused spectral images, and the identification of tumor lesion regions, specifically includes: Based on the fused spectral image, the trained target detection model is used to detect tumor targets, identify tumor lesion areas, and mark the tumor lesion areas with target detection anchor boxes. The target detection model uses the YOLOv3 target detection algorithm and is trained using fused spectral images of several manually annotated tumor lesion regions as a training sample set to obtain the trained target detection model.

5. The intraoperative rapid tumor classification system based on hyperspectral technology as described in claim 1, characterized in that, The process of sampling the identified tumor lesion region using multiple pixels to obtain the frequency characteristics of each preferred frequency at each sampling point is specifically as follows: The tumor lesion area marked by the target detection anchor frame is determined. The center of the anchor frame is taken as the reference point. The reference point is taken as point 0, the midpoint from point 0 to the boundary of the anchor frame is taken as points 1-4, and the midpoint from points 1-4 to point 0 is taken as points 5-8. Multi-pixel sampling is performed on these 8 points 1-8 to obtain the frequency characteristics of each sampling point at each preferred frequency.

6. The intraoperative rapid tumor classification system based on hyperspectral technology as described in claim 1, characterized in that, The benign tumor identification model uses a convolutional neural network connecting a fully connected layer and a binary classifier to distinguish benign tumors from others. The training process of the benign tumor identification model is as follows: using the one-dimensional frequency characteristic data of multiple sampling points of the hyperspectral image of the benign tumor lesion area as the training sample dataset, the model is trained to obtain the identification model for benign tumors or other benign tumors.

7. The intraoperative rapid tumor classification system based on hyperspectral technology as described in claim 1, characterized in that, The malignant tumor identification model uses a convolutional neural network and a binary classifier to distinguish between malignant tumors and others; The training process of the malignant tumor identification model is as follows: using the one-dimensional frequency characteristic data of multiple sampling points of the hyperspectral image of the malignant tumor lesion area as the training sample dataset, training is performed to obtain an identification model for malignant tumors or other malignant tumors.

8. The intraoperative rapid tumor classification system based on hyperspectral technology as described in claim 1, characterized in that, The method of identifying and determining the type of tumor lesion based on the benign probability matrix and the malignant probability matrix is ​​as follows: Based on the benign and malignant probability matrices, a probability matrix analysis is performed: If the benign probability matrix is ​​benign and the malignant probability matrix is ​​other, then the final tumor lesion type is determined to be a benign tumor. If the benign probability matrix is ​​"other" and the malignant probability matrix is ​​"malignant", then the final tumor lesion type is determined to be a malignant tumor. If the benign probability matrix is ​​"other" and the malignant probability matrix is ​​also "other", then the identified tumor lesion area is determined to be normal tissue or other non-tumor lesions. If the benign probability matrix is ​​benign and the malignant probability matrix is ​​malignant, further processing is still required.

9. The intraoperative rapid tumor classification system based on hyperspectral technology as described in claim 8, characterized in that, The subsequent processing involves inputting the one-dimensional frequency characteristic data of the identified tumor lesion region into the trained classification model to determine the type of the tumor lesion.

Images