A Low-Cost Intelligent Inference Method and Device for Nasopharyngeal Carcinoma Medical Images

Abstract

The present invention provides a low-cost intelligent inference method and device for nasopharyngeal carcinoma medical images, which relates to the technical fields of image processing and artificial intelligence. The method includes: collecting nasopharyngeal carcinoma medical images; obtaining a local image set by using a shearing and cropping method for the nasopharyngeal carcinoma medical images; sending the local image set into an artificial intelligence model for inference to obtain a segmentation result of the local image set; and cropping the core region of the segmentation result of the local image set and stitching them together to form a final segmentation result. By adopting the present invention, the core region cropping method is separated from the network design process, the cropping operation in the model design process is removed, so that the output size of the network model is the same as the input size, and the core region cropping is used as a post-processing operation with adjustable hyperparameters. This method further reduces the limitations of the input images and expands the adaptation ability of the deep model to different hardware resources.

Description

Technical Field

[0001] The present invention relates to the technical fields of image processing and artificial intelligence, and particularly to a low-cost intelligent inference method and device for nasopharyngeal carcinoma medical images. Background Art

[0002] In nasopharyngeal clinical tasks, by introducing image segmentation technology based on deep learning, the tumor regions in CT (Computed Tomography), PET (Positron Emission Computed Tomography), MRI (Magnetic Resonance Imaging), and pathological images can be intelligently extracted, providing technical and data support for doctors.

[0003] In actual inference applications, since the images to be analyzed usually have the characteristics of high resolution and large size, the image segmentation model based on deep learning requires a large amount of video memory resources during calculation, and simple image cropping for image segmentation will be affected by the zero-padding operation of convolution, resulting in edge recognition errors. Therefore, there is an urgent need to design a low-cost intelligent inference method for nasopharyngeal carcinoma medical images.

[0004] In Unet (Ronneberger O, Fischer P, Brox T. U-net: Convolutional networks for biomedical image segmentation [C] / / International Conference on Medical image computing and computer-assisted intervention. Springer, Cham, 2015: 234-241), to address the problem of excessive occupation of video memory resources in inference applications, an Overlap-tile shear cutting strategy is proposed to alleviate the error phenomenon at the edges during local image inference caused by the zero-padding operation in convolution. However, this strategy solidifies the core region cropping method in the network structure design, making the output size of the network model smaller than the input size. This strategy requires the input image size to be not less than the size during network structure cropping, thereby restricting the sizes of the input image and the core region. While lacking flexibility, it limits the minimum video memory occupation during network inference and ultimately restricts the adaptability of the network model to different hardware resources.

[0005] In the existing analysis tasks of nasopharyngeal carcinoma medical images, due to the different sizes of images collected by different machines with different parameters and modalities, it poses a challenge to the flexibility of model design. Therefore, there is an urgent need for a flexible and low-cost intelligent inference method for nasopharyngeal carcinoma medical images to reduce the video memory cost of the intelligent model and the errors caused by cropped images. Summary of the Invention

[0006] Aiming at the problems in the prior art, such as the lack of flexibility in model design, high video memory cost of the intelligent model, and easy errors caused by cropped images, the present invention proposes a low-cost intelligent inference method and device for nasopharyngeal carcinoma medical images.

[0007] To solve the above technical problems, the present invention provides the following technical solutions:

[0008] On the one hand, a low-cost intelligent inference method for nasopharyngeal carcinoma medical images is provided. This method is applied to an electronic device and includes:

[0009] S1: Collect nasopharyngeal carcinoma medical images;

[0010] S2: Preprocess the nasopharyngeal carcinoma medical images;

[0011] S3: Through the shear cutting method, locally crop the preprocessed nasopharyngeal carcinoma medical images to obtain a local image set;

[0012] S4: Input the local image set into a preset artificial intelligence model for inference and output the segmentation result of the local image set;

[0013] S5: Crop the core area of the segmentation result of the local image set and splice it into the final segmentation result to complete the low-cost intelligent inference of nasopharyngeal carcinoma medical images.

[0014] Optionally, the nasopharyngeal carcinoma medical images include one or more of nasopharyngoscopy imaging, head computed tomography (CT) images, head positron emission tomography (PET) images, head magnetic resonance imaging (MRI) images, and tumor pathological section images.

[0015] Optionally, in step S2, preprocessing the nasopharyngeal carcinoma medical images includes:

[0016] S21: Set the width w of the nasopharyngeal carcinoma medical image d , the cropping width w t , the core area width w c , the interval width w i ; w t , w d and w i satisfy the relationship of the following formula (1):

[0017] wt = w c + 2w i (1)

[0018] where w t , w c , w i ≥ 0; w t ≤ w d ;

[0019] S22: By using the Pad filling method, the nasopharyngeal carcinoma medical images are amplified by w i pixels in the up, down, left, and right directions respectively to form the nasopharyngeal carcinoma medical images after amplified pixels.

[0020] Optionally, the Pad filling method includes one or more of zero padding and replication padding.

[0021] Optionally, in step S3, by using the shearing and cropping method, the preprocessed nasopharyngeal carcinoma medical images are locally cropped to obtain a set of local images, including:

[0022] By using the shearing and cropping method, the nasopharyngeal carcinoma medical images after amplified pixels are locally cropped to obtain a set of local images; the length and width dimensions of all local images in the cropped set of local images are w t .

[0023] Optionally, locally cropping the nasopharyngeal carcinoma medical images after amplified pixels includes:

[0024] Let (i, j) be the local image cropped in the i-th row and j-th column in the original image, and its upper left coordinate is ((i - 1) × (w i + w c ), (j - 1) × (w i + w c )); when j takes the maximum value, the upper left coordinate of its local image is ((i - 1) × (w i + w c ), w d - w t ); when i takes the maximum value, the upper left coordinate of its local image is (w d - w t , (j - 1) × (w i + w c ))).

[0025] Optionally, in step S4, the artificial intelligence model is an image segmentation model based on a convolutional neural network.

[0026] Optionally, in step S5, the core region of the segmentation result of the local image set is cropped and stitched together to form the final segmentation result, completing the low-cost intelligent inference of nasopharyngeal carcinoma medical images, including:

[0027] For the segmentation result of each local image, crop the region [w i :w i +w c ,w i :w i +w c and directly stitch it together in the original cropping order to form a segmentation result of the same size as the original nasopharyngeal carcinoma medical image, completing the low-cost intelligent inference of nasopharyngeal carcinoma medical images.

[0028] On the one hand, a low-cost intelligent inference device for nasopharyngeal carcinoma medical images is provided. The device is applied to an electronic device and includes:

[0029] An image acquisition module for acquiring nasopharyngeal carcinoma medical images;

[0030] A preprocessing module for preprocessing nasopharyngeal carcinoma medical images;

[0031] A shearing and cropping module for locally cropping the preprocessed nasopharyngeal carcinoma medical image by the shearing and cropping method to obtain a local image set;

[0032] A model inference module for inputting the local image set into a preset artificial intelligence model for inference and outputting the segmentation result of the local image set;

[0033] A stitching module for cropping the core region of the segmentation result of the local image set and stitching it together to form the final segmentation result, completing the low-cost intelligent inference of nasopharyngeal carcinoma medical images.

[0034] Optionally, the nasopharyngeal carcinoma medical images include one or more of nasopharyngoscopy imaging, head computed tomography (CT) images, head positron emission tomography (PET) images, head magnetic resonance imaging (MRI) images, and tumor pathological section images.

[0035] On the one hand, an electronic device is provided. The electronic device includes a processor and a memory, and at least one instruction is stored in the memory. The at least one instruction is loaded and executed by the processor to implement the above-mentioned low-cost intelligent inference method for nasopharyngeal carcinoma medical images.

[0036] On the one hand, a computer-readable storage medium is provided. At least one instruction is stored in the storage medium. The at least one instruction is loaded and executed by a processor to implement the above-mentioned low-cost intelligent inference method for nasopharyngeal carcinoma medical images.

[0037] The above technical solutions of the embodiments of the present invention have at least the following beneficial effects:

[0038] In the above solution, the method provided by the present invention separates the core region cropping method in the image inference process from the network design process, removes the cropping operation in the model design process, makes the output size of the network model the same as the input size, and takes the core region cropping as a post-processing operation that can adjust hyperparameters. This method further reduces the limitations of the input image and expands the adaptation ability of the deep model to different hardware resources; it solves the problems of high video memory cost of the intelligent model and easy errors caused by cropped images. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] In order to more clearly illustrate the technical solutions in the embodiments of the present invention, the following will briefly introduce the drawings required for the description of the embodiments. Obviously, the following drawings are only some embodiments of the present invention. For those of ordinary skill in the art, without creative efforts, other drawings can be obtained based on these drawings.

[0040] Figure 1 is a flowchart of a low-cost intelligent inference method for pharyngeal cancer medical images provided by an embodiment of the present invention;

[0041] Figure 2 is a flowchart of a low-cost intelligent inference method for pharyngeal cancer medical images provided by an embodiment of the present invention;

[0042] Figure 3 is a schematic diagram of cropping for nasopharyngeal MRI image segmentation of a low-cost intelligent inference method for pharyngeal cancer medical images provided by an embodiment of the present invention;

[0043] Figure 4 is a block diagram of a low-cost intelligent inference device for pharyngeal cancer medical images provided by an embodiment of the present invention;

[0044] Figure 5 is a schematic diagram of the structure of an electronic device provided by an embodiment of the present invention. DETAILED DESCRIPTION OF THE EMBODIMENTS

[0045] To make the technical problems, technical solutions and advantages to be solved by the present invention clearer, the following will be described in detail with reference to the drawings and specific embodiments.

[0046] An embodiment of the present invention provides a low-cost intelligent inference method for pharyngeal cancer medical images. This method can be implemented by an electronic device, which can be a terminal or a server. As Figure 1 shown in the flowchart of the low-cost intelligent inference method for pharyngeal cancer medical images, the processing flow of this method can include the following steps:

[0047] S101: Collect nasopharyngeal carcinoma medical images;

[0048] S102: Preprocess the nasopharyngeal carcinoma medical images;

[0049] S103: Use the shearing and cropping method to locally crop the preprocessed nasopharyngeal carcinoma medical images to obtain a set of local images;

[0050] S104: Input the set of local images into a preset artificial intelligence model for inference and output the segmentation results of the set of local images;

[0051] S105: Crop the core regions of the segmentation results of the set of local images, piece them together to form the final segmentation results, and complete the low-cost intelligent inference of nasopharyngeal carcinoma medical images.

[0052] Optionally, the nasopharyngeal carcinoma medical images include one or more of: nasopharyngoscope imaging, head computed tomography (CT) images, head positron emission tomography (PET) images, head magnetic resonance imaging (MRI) images, and tumor pathological section images.

[0053] Optionally, in step S102, preprocessing the nasopharyngeal carcinoma medical images includes:

[0054] S121: Set the width w of the nasopharyngeal carcinoma medical images d , the cropping width w t , the core region width w c , the interval width w i ; w t , w d and w i satisfy the relationship as shown in the following formula (1):

[0055] w t = w c + 2w i (1)

[0056] where, w t , w c , w i ≥ 0; w t ≤ w d ;

[0057] S122: Use the Pad filling method to expand the nasopharyngeal carcinoma medical images by w i pixels in the up, down, left, and right directions respectively to form the nasopharyngeal carcinoma medical images after expanding the pixels.

[0058] Optionally, the Pad filling method includes one or more of zero filling and replication filling.

[0059] Optionally, in step S103, local image cropping is performed on the preprocessed nasopharyngeal carcinoma medical image by the shear cutting method to obtain a local image set, including:

[0060] Local image cropping is performed on the nasopharyngeal carcinoma medical image after pixel amplification by the shear cutting method to obtain a local image set; the length and width dimensions of all local images in the cropped local image set are both w t .

[0061] Optionally, local image cropping of the nasopharyngeal carcinoma medical image after pixel amplification includes:

[0062] Let (i, j) be the local image cropped in the i-th row and j-th column in the original image, and its upper left corner coordinates are ((i - 1) × (w i +w c ),(j - 1) × (w i +w c )); when j takes the maximum value, the upper left corner coordinates of its local image are ((i - 1) × (w i +w c ), w d -w t ); when i takes the maximum value, the upper left corner coordinates of its local image are (w d -w t ,(j - 1) × (w i +w c ))

[0063] Optionally, in step S104, the artificial intelligence model is an image segmentation model based on a convolutional neural network.

[0064] Optionally, in step S105, cropping the core region of the segmentation result of the local image set and stitching it into the final segmentation result to complete the low-cost intelligent inference of the nasopharyngeal carcinoma medical image, including:

[0065] For the segmentation result of each local image, crop [w i :w i +w c , w i :w i +w c and directly stitch it into a segmentation result of the same size as the original nasopharyngeal carcinoma medical image according to the original cropping order to complete the low-cost intelligent inference of the nasopharyngeal carcinoma medical image.

[0066] In the embodiments of the present invention, the core region cropping method in the image inference process is separated from the network design process by the method provided by the present invention, and the cropping operation in the model design process is removed, so that the output size of the network model is the same as the input size. And the core region cropping is used as a post-processing operation that can adjust hyperparameters. This method further reduces the limitations of the input image and enhances the adaptation ability of the deep model to different hardware resources.

[0067] Embodiments of the present invention provide a low-cost intelligent inference method for pharyngeal cancer medical images. This method can be implemented by an electronic device, which can be a terminal or a server. As Figure 2 shown in the flowchart of the low-cost intelligent inference method for pharyngeal cancer medical images, the processing flow of this method can include the following steps:

[0068] S201: Collect pharyngeal cancer medical images.

[0069] In a feasible implementation, the pharyngeal cancer medical images may include one or more of nasopharyngoscopy imaging, head computed tomography (CT) images, head positron emission tomography (PET) images, head magnetic resonance imaging (MRI) images, and tumor pathological section images.

[0070] S202: Set the width w of the pharyngeal cancer medical image d , cropping width w t , core region width w c , interval width w i ; w t , w d and w i satisfy the relationship of the following formula (1):

[0071] w t = w c + 2w i (1)

[0072] Wherein, w t , w c , w i ≥ 0; w t ≤ w d ;

[0073] S203: Through the padding method, expand the pharyngeal cancer medical image by w i pixels in the four directions of up, down, left, and right to form an expanded pharyngeal cancer medical image with expanded pixels.

[0074] In a feasible implementation, the padding method includes one or more of zero padding and copy padding.

[0075] S204: Crop the preprocessed nasopharyngeal carcinoma medical image by the shearing and cropping method to obtain a set of local images.

[0076] In a feasible implementation, as Figure 3 shown, crop the local images of the nasopharyngeal carcinoma medical image after pixel amplification by the shearing and cropping method to obtain a set of local images; the length and width of all local images in the cropped set of local images are both w t .

[0077] In a feasible implementation, let (i, j) be the local image cropped in the i-th row and j-th column in the original image, and its upper left corner coordinates are ((i - 1) × (w i +w c ),(j - 1) × (w i +w c )); when j takes the maximum value, the upper left corner coordinates of its local image are ((i - 1) × (w i +w c ), w d -w t ); when i takes the maximum value, the upper left corner coordinates of its local image are (w d -w t ,(j - 1) × (w i +w c ).

[0078] S205: Input the set of local images into a preset artificial intelligence model for inference, and output the segmentation result of the set of local images.

[0079] In a feasible implementation, the artificial intelligence model is an image segmentation model based on a convolutional neural network.

[0080] S206: Crop the core area of the segmentation result of the set of local images, and splice it into the final segmentation result to complete the low-cost intelligent inference of the nasopharyngeal carcinoma medical image.

[0081] In a feasible implementation, for the segmentation result of each local image, crop the area [w i : w i +w c , w i : w i +w c and splice it directly in the original cropping order into a segmentation result of the same size as the original nasopharyngeal carcinoma medical image to complete the low-cost intelligent inference of the nasopharyngeal carcinoma medical image.

[0082] The above method will be described in detail below through the test results corresponding to an embodiment of the present invention:

[0083] In this embodiment, 36 large-size images with a size of 2800×1600 are selected for analysis, and the average value of the information change amount (VI) is taken as the analysis result.

[0084] (1) w t and w i Effect on performance

[0085] Compared with the shearing and cropping strategy described in the classical Unet, the method proposed in the present invention can modify hyperparameters during the inference process to adapt to different hardware configurations. Therefore, the present invention first analyzes the influence of hyperparameters on the method proposed herein.

[0086] w t and w i are hyperparameters of the present invention. The present invention explores the performance of different input parameters under the same hardware resource conditions (a single NVIDIA Tesla V100). As shown in Table 1, the effects on the maximum video memory occupancy, time consumption, and VI are respectively shown under different inputs of w t and w i . Among them, the input image size needs to be a multiple of 2 s .

[0087] Table 1 Effects of changes in w t and w i on performance

[0088]

[0089] It can be seen from Table 1 that the video memory occupancy during inference increases with the increase of w t . During the model inference process, the video memory occupancy mainly includes the video memory occupied by network parameters and the feature maps of the input images. When the input image is small, the video memory is mainly occupied by network parameters (118.42MB). Therefore, when w t increases from 16 to 64, the video memory occupancy does not increase significantly.

[0090] The time consumption of model inference decreases with the increase of the input image size. Its time consumption mainly includes the transmission time between the CPU (central processing units) and the GPU (graphics processing unit) and the GPU parallel computing time. When the input image is small, there are many slices, the transmission time is large, and the GPU parallel time is small; when the input image is large, there are few slices, the transmission time is small, and the GPU parallel time is large; but the transmission time increases linearly, and the speed increase ratio of GPU parallel computing is larger when the image is larger. Therefore, the time consumption of model inference decreases with the increase of the input image size;

[0091] Moreover, as the size of the input image increases, the effective receptive field during model inference increases, so the overall VI error rate decreases. Although the method proposed in the present invention can reduce the size of the input image to 16 during forward inference, compared with the classical Unet, which can only reduce the size of the input image to a minimum of 392 during fixed cropping and requires a minimum size of 140 during adaptive cropping, the method proposed in the present invention also presents acceptable performance results when the input image sizes are 32, 64, and 128, indicating that the method proposed in the present invention can save more video memory occupancy.

[0092] This embodiment further explores the influence of changing w t while keeping w i unchanged on the result. For the convenience of analysis, this embodiment sets w t = 128, and the results are shown in Table 2.

[0093] Table 2 Influence of the change of w i on performance

[0094]

[0095] As can be seen from Table 2, as w i becomes larger, the core area becomes smaller, the influence of the zero-padding operation on the result during model calculation becomes smaller, and the VI error rate becomes lower; the larger w i is, the smaller the core area is, the more slices there are, and the more time-consuming it is; considering both the time and VI metrics, and taking into account that in practical applications, medical staff can choose a machine with a lower hardware configuration to run the present invention, finally w i is selected to be 16.

[0096] (2) Evaluation results under different hardware machines

[0097] Table 3 Evaluation results under different hardware resources

[0098]

[0099] This embodiment further compares the influence of selecting different hyperparameters on speed and performance under different hardware resources. This embodiment selects a high-performance GPU computing workstation and a low-performance personal laptop for comparative analysis to prove that the strategy proposed in this embodiment can enable the deep network model to be deployed on any model, where the low-performance personal laptop mainly relies on memory for calculation and analysis. As shown in Table 3. It can be seen from the table that the algorithm of this embodiment can run on both the high-performance workstation and the low-performance personal laptop, and the larger w t is, the shorter the time-consuming is, and the lower the VI error amount is.

[0100] Figure 4 is a block diagram of a low-cost intelligent inference device for nasopharyngeal carcinoma medical images shown according to an exemplary embodiment. Refer to Figure 4, the device 300 includes:

[0101] An image acquisition module 310 for acquiring nasopharyngeal carcinoma medical images;

[0102] A preprocessing module 320 for preprocessing nasopharyngeal carcinoma medical images;

[0103] A shearing and cropping module 330 for locally cropping the preprocessed nasopharyngeal carcinoma medical images by the shearing and cropping method to obtain a local image set;

[0104] A model inference module 340 for inputting the local image set into a preset artificial intelligence model for inference and outputting the segmentation result of the local image set;

[0105] A stitching module 350 for cropping the core regions of the segmentation results of the local image set and stitching them into a final segmentation result to complete the low-cost intelligent inference of nasopharyngeal carcinoma medical images.

[0106] Optionally, the nasopharyngeal carcinoma medical images include one or more of nasopharyngoscope imaging, head CT images, head PET images, head MRI images, and tumor pathological section images.

[0107] Optionally, the preprocessing module 320 is further configured to set the width w of the nasopharyngeal carcinoma medical image d , the cropping width w t , the core region width w c , the interval width w i ; w t , w d and w i satisfy the relationship of the following formula (1):

[0108] w t = w c + 2w i (1)

[0109] Wherein, w t , w c , w i ≥ 0; w t ≤ w d ;

[0110] By the padding method, the nasopharyngeal carcinoma medical image is amplified by w i pixels in the up, down, left, and right directions respectively to form an amplified nasopharyngeal carcinoma medical image with amplified pixels.

[0111] Optionally, the padding method includes one or more of zero padding and replication padding.

[0112] Optionally, the shearing and cropping module 330 is further configured to perform local image cropping on the nasopharyngeal carcinoma medical image after amplifying pixels by means of shearing and cropping to obtain a set of local images; the lengths and widths of all local images in the cropped set of local images are both w t .

[0113] Optionally, the shearing and cropping module 330 is further configured to set (i, j) as the local image of the i-th row and j-th column cropped in the original image, and its upper left coordinate is ((i - 1) × (w i +w c ),(j - 1) × (w i +w c )); when j takes the maximum value, the upper left coordinate of its local image is ((i - 1) × (w i +w c ), w d -w t ); when i takes the maximum value, the upper left coordinate of its local image is (w d -w t ,(j - 1) × (w i +w c ).

[0114] Optionally, the artificial intelligence model is an image segmentation model based on a convolutional neural network.

[0115] Optionally, the stitching module 350 is further configured to, for the segmentation result of each local image, crop [w i :w i +w c , w i :w i +w c and directly stitch them in the original cropping order into a segmentation result of the same size as the original nasopharyngeal carcinoma medical image to complete low-cost intelligent inference of the nasopharyngeal carcinoma medical image.

[0116] In the embodiments of the present invention, the method provided by the present invention separates the core region cropping method in the image inference process from the network design process, removes the cropping operation in the model design process, makes the output size of the network model the same as the input size, and takes the core region cropping as a post-processing operation that can adjust hyperparameters. This method further reduces the limitation of the input image and amplifies the adaptation ability of the deep model to different hardware resources.

[0117] Figure 5FIG. 0 is a schematic structural diagram of an electronic device 400 provided by an embodiment of the present invention. The electronic device 400 may vary greatly due to different configurations or performances, and may include one or more central processing units (CPUs) 401 and one or more memories 402. Among them, at least one instruction is stored in the memory 402, and the at least one instruction is loaded and executed by the processor 401 to implement the steps of the following low-cost intelligent inference method for nasopharyngeal carcinoma medical images:

[0118] S1: Collect nasopharyngeal carcinoma medical images;

[0119] S2: Preprocess the nasopharyngeal carcinoma medical images;

[0120] S3: Through the shearing and cropping method, locally crop the preprocessed nasopharyngeal carcinoma medical images to obtain a local image set;

[0121] S4: Input the local image set into a preset artificial intelligence model for inference, and output the segmentation result of the local image set;

[0122] S5: Crop the core area of the segmentation result of the local image set and splice it into the final segmentation result to complete the low-cost intelligent inference of nasopharyngeal carcinoma medical images.

[0123] In an exemplary embodiment, a computer-readable storage medium is also provided, such as a memory including instructions, and the above instructions can be executed by a processor in a terminal to complete the above low-cost intelligent inference method for nasopharyngeal carcinoma medical images. For example, the computer-readable storage medium may be a ROM, a random access memory (RAM), a CD-ROM, a magnetic tape, a floppy disk, an optical data storage device, etc.

[0124] Those of ordinary skill in the art can understand that all or part of the steps of implementing the above embodiments can be completed by hardware, or can be completed by a program instructing related hardware. The program can be stored in a computer-readable storage medium, and the above-mentioned storage medium can be a read-only memory, a magnetic disk, or an optical disc, etc.

[0125] The above are only the preferred embodiments of the present invention, and are not intended to limit the present invention. Any modifications, equivalent replacements, improvements, etc. made within the spirit and principles of the present invention shall be included in the protection scope of the present invention.

Claims

1. A low-cost intelligent inference method for nasopharyngeal carcinoma medical images, characterized in that, Including: S1: Collect nasopharyngeal carcinoma medical images; S2: Preprocess the nasopharyngeal carcinoma medical images; In the step S2, preprocessing the nasopharyngeal carcinoma medical images includes: S21: Set the width w of the nasopharyngeal carcinoma medical image d , the cropping width w t , the width w of the core area c , the interval width w i ; The w t , w d and w i satisfy the relationship of the following formula (1): w t = w c + 2w i (1) Among them, w t , w c , w i ≥0; w t ≤w d ; S22: By using the Pad filling method, the nasopharyngeal carcinoma medical image is amplified by w pixels in each of the four directions of up, down, left, and right to form a nasopharyngeal carcinoma medical image with amplified pixels; i Pixels are added to form a nasopharyngeal carcinoma medical image with amplified pixels; S3: Through the shearing and cropping method, locally crop the preprocessed nasopharyngeal carcinoma medical images to obtain a local image set; In the step S3, through the shearing and cropping method, locally crop the preprocessed nasopharyngeal carcinoma medical images to obtain a local image set, including: Through the shearing and cropping method, local image cropping is performed on the nasopharyngeal carcinoma medical image after amplifying pixels to obtain a local image set; the length and width dimensions of all local images in the cropped local image set are both w t ; Locally crop the nasopharyngeal carcinoma medical images after pixel amplification, including: Let (i, j) be the local image cropped from the original image in the i-th row and j-th column, and its upper-left corner coordinates are ((i - 1) × (w i + w c ), (j - 1) × (w i + w c )); when j takes the maximum value, the upper-left corner coordinates of its local image are ((i - 1) × (w i + w c ), w d - w t ); when i takes the maximum value, the upper-left corner coordinates of its local image are (w d - w t , (j - 1) × (w i + w c )); S4: Input the local image set into a preset artificial intelligence model for inference, and output the segmentation result of the local image set; S5: Crop the core region of the segmentation result of the local image set, and splice it into the final segmentation result to complete the low-cost intelligent inference of nasopharyngeal carcinoma medical images.

2. The low-cost intelligent inference method for nasopharyngeal carcinoma medical images according to claim 1, characterized in that The nasopharyngeal carcinoma medical images include one or more of nasopharyngoscope imaging, head computed tomography (CT) images, head positron emission tomography (PET) images, head magnetic resonance imaging (MRI) images, and tumor pathological section images.

3. The low-cost intelligent inference method for nasopharyngeal carcinoma medical images according to claim 2, wherein The padding Pad method includes one or more of zero padding and replication padding.

4. The low-cost intelligent inference method for nasopharyngeal carcinoma medical images according to claim 1, wherein In the step S4, the artificial intelligence model is an image segmentation model based on a convolutional neural network.

5. The low-cost intelligent inference method for nasopharyngeal carcinoma medical images according to claim 4, wherein In the step S5, crop the core region of the segmentation result of the local image set, and splice it into the final segmentation result to complete the low-cost intelligent inference of nasopharyngeal carcinoma medical images, including: For the segmentation result of each of the local images, crop the part where [w i :w i +w c ,w i :w i +w c and directly stitch them together in the original cropping order to form a segmentation result of the same size as the original nasopharyngeal carcinoma medical image, thus completing the low-cost intelligent inference of nasopharyngeal carcinoma medical images.

6. A low-cost intelligent inference device for nasopharyngeal carcinoma medical images, characterized in that, The device is applicable to the low-cost intelligent inference method of nasopharyngeal carcinoma medical images according to any one of the above claims 1-5. The device includes: An image acquisition module for collecting nasopharyngeal carcinoma medical images; A preprocessing module for preprocessing the nasopharyngeal carcinoma medical images; A shearing and cropping module for locally cropping the preprocessed nasopharyngeal carcinoma medical images through the shearing and cropping method to obtain a local image set; A model inference module for inputting the local image set into an artificial intelligence model for inference and outputting the segmentation result of the local image set; A splicing module for cropping the core region of the segmentation result of the local image set and splicing it into the final segmentation result to complete the low-cost intelligent inference of nasopharyngeal carcinoma medical images.

7. The low-cost intelligent inference device for nasopharyngeal carcinoma medical images according to claim 6, wherein The nasopharyngeal carcinoma medical images include one or more of nasopharyngoscope imaging, head computed tomography (CT) images, head positron emission tomography (PET) images, head magnetic resonance imaging (MRI) images, and tumor pathological section images.

Images