Intelligent sickbed control system based on femur detection

Abstract

The embodiment of the invention discloses an intelligent sickbed control system based on femur detection. A specific embodiment of the system comprises an X-ray machine configured to acquire femur image data; a data processor configured to execute the following steps: performing image detection processing on femur image data received from the X-ray machine; performing gray mapping processing on the femur image data; obtaining standardized femur image data; performing instance segmentation processing on the standardized femur image data; carrying out key point detection processing on the femoral ash degree image data to obtain coordinate data of each key point; generating a thighbone detection result; and the intelligent sickbed controller is configured to control the intelligent sickbed to execute a position adjustment task based on the thighbone detection result. According to the embodiment, the frequency of occurrence of system jamming or no response can be reduced.

Description

Technical Field

[0001] The embodiments disclosed herein relate to the field of computer technology, and more specifically to an intelligent hospital bed control system based on femoral bone detection. Background Technology

[0002] The intelligent hospital bed control system based on femoral bone detection involves detecting the user's femur and then controlling the intelligent hospital bed based on the detection results to adjust the user's position. Currently, the common approach for controlling intelligent hospital beds based on femoral bone detection is as follows: first, acquire femoral bone image data; then, process the femoral bone image data directly using an end-to-end neural network model, outputting the corresponding positional information; and finally, based on this positional information, control the intelligent hospital bed to adjust the user's position.

[0003] However, when using the above methods for intelligent hospital bed control, the following technical problems often arise: The system response time is long. When directly processing femoral image data end-to-end through the model, the computational complexity of the model is high, which can easily lead to excessive computational load on the system when running the model, causing system freezes and resulting in a long system response time. Summary of the Invention

[0004] The summary portion of this disclosure is intended to provide a brief overview of the concepts, which will be described in detail in the detailed description portion. This summary portion is not intended to identify key or essential features of the claimed technical solutions, nor is it intended to limit the scope of the claimed technical solutions.

[0005] Some embodiments of this disclosure propose an intelligent bed control system based on femoral bone detection to solve one or more of the technical problems mentioned in the background section above.

[0006] In a first aspect, some embodiments of this disclosure provide an intelligent hospital bed control system based on femoral detection. The system includes: an X-ray machine configured to acquire femoral image data; and a data processor configured to perform the following steps: performing image detection processing on the femoral image data received from the X-ray machine to obtain an image detection result; in response to determining that the image detection result meets preset image detection conditions, performing grayscale mapping processing on the femoral image data based on a preset historical femoral image dataset to obtain femoral grayscale image data; performing image standardization processing on the femoral grayscale image data to obtain standardized femoral image data; performing instance segmentation processing on the standardized femoral image data to obtain individual femoral mask data; performing keypoint detection processing on the femoral grayscale image data to obtain individual keypoint coordinate data, wherein each keypoint coordinate data corresponds to a keypoint name; generating a femoral detection result based on the individual femoral mask data, the individual keypoint coordinate data, and the keypoint names corresponding to the individual keypoint coordinate data; and an intelligent hospital bed controller configured to control the intelligent hospital bed to perform a position adjustment task based on the femoral detection result.

[0007] The above embodiments of this disclosure have the following beneficial effects: the intelligent bed control system based on femoral detection in some embodiments of this disclosure can reduce the system response time. Specifically, the reason for the long system response time is that when the femoral image data is processed end-to-end directly through the model, the computational complexity of the model is high, which can easily lead to excessive computational load on the system when running the model, causing the system to freeze and thus resulting in a long system response time. Based on this, in the intelligent bed control system based on femoral detection in some embodiments of this disclosure, firstly, an X-ray machine is configured to acquire femoral image data. Thus, femoral image data can be obtained. Then, a data processor is configured to perform the following steps: perform image detection processing on the femoral image data received from the X-ray machine to obtain an image detection result. Thus, the image can be detected. Secondly, in response to determining that the above image detection result meets the preset image detection conditions, grayscale mapping processing is performed on the above femoral image data based on a preset historical femoral image dataset to obtain femoral grayscale image data. Thus, the image is processed only when the detected image meets the processing conditions to obtain femoral grayscale image data. Then, the aforementioned femoral grayscale image data is subjected to image standardization processing to obtain standardized femoral image data. Next, instance segmentation processing is performed on the standardized femoral image data to obtain individual femoral mask data. Then, keypoint detection processing is performed on the aforementioned femoral grayscale image data to obtain coordinate data for each keypoint, where each keypoint coordinate corresponds to a keypoint name. Then, keypoint recognition is performed on the femoral grayscale image data to obtain coordinate data for each keypoint. Then, based on the aforementioned femoral mask data, the aforementioned keypoint coordinate data, and the corresponding keypoint names, a femoral detection result is generated. Finally, the intelligent bed controller is configured to control the intelligent bed to perform a patient positioning adjustment task based on the aforementioned femoral detection result. Therefore, the intelligent hospital bed can be controlled based on the generated femoral detection results to adjust the user's position. Because it first generates femoral mask data and key point coordinate data corresponding to the image, then uses this data to generate femoral detection results, and finally controls the intelligent hospital bed to perform position adjustment tasks based on these results—instead of directly generating position information corresponding to femoral image data from a large end-to-end model—the computational load on the system can be reduced. This decreases the frequency of system crashes due to excessive computational load and reduces system response time. Attached Figure Description

[0008] The above and other features, advantages, and aspects of the embodiments of this disclosure will become more apparent from the accompanying drawings and the following detailed description. Throughout the drawings, the same or similar reference numerals denote the same or similar elements. It should be understood that the drawings are schematic, and elements are not necessarily drawn to scale.

[0009] Figure 1 This is a schematic diagram of the structure of an electronic device suitable for implementing some embodiments of the present disclosure; Figure 2 It can be a schematic diagram showing the location of each key point in the femoral image data, corresponding to each key point name; Figure 3 This can be a schematic diagram of the head screw portion of an intramedullary nail; Figure 4 It can be used to visualize images after segmenting intramedullary nails and femoral medullary cavities. Detailed Implementation

[0010] Embodiments of this disclosure will now be described in more detail with reference to the accompanying drawings. While some embodiments of this disclosure are shown in the drawings, it should be understood that this disclosure can be implemented in various forms and should not be construed as limited to the embodiments set forth herein. Rather, these embodiments are provided to provide a more thorough and complete understanding of this disclosure. It should be understood that the accompanying drawings and embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of protection of this disclosure.

[0011] It should also be noted that, for ease of description, only the parts relevant to the invention are shown in the accompanying drawings. Unless otherwise specified, the embodiments and features described in this disclosure can be combined with each other.

[0012] It should be noted that the concepts of "first" and "second" mentioned in this disclosure are used only to distinguish different devices, modules or units, and are not used to limit the order of functions performed by these devices, modules or units or their interdependencies.

[0013] It should be noted that the terms "a" and "a plurality of" used in this disclosure are illustrative rather than restrictive, and those skilled in the art should understand that, unless otherwise expressly indicated in the context, they should be understood as "one or more".

[0014] The names of messages or information exchanged between multiple devices in the embodiments of this disclosure are for illustrative purposes only and are not intended to limit the scope of such messages or information.

[0015] This disclosure will now be described in detail with reference to the accompanying drawings and embodiments.

[0016] Figure 1A schematic diagram of the structure of some embodiments of the intelligent bed control system based on femoral detection according to the present disclosure is shown. The intelligent bed control system based on femoral detection includes: an X-ray machine 1, a data processor 2, and an intelligent bed controller 3.

[0017] The data processor 2 can be a device capable of processing data. For example, the data processor 2 can be a server. The smart bed controller 3 can be a device capable of adjusting the smart bed by sending commands. For example, the smart bed controller 3 can be a server.

[0018] In some embodiments, the X-ray machine 1 can be configured to acquire femoral image data. This femoral image data can be an X-ray film of the femur of a target user. The target user can be a user who has undergone femoral fracture surgery and has had an intramedullary nail implanted in their femur. The femoral image data corresponds to user information. This user information can be information corresponding to the target user. This user information may include, but is not limited to: user name, user gender, user ID number, and user bed number. In practice, the X-ray machine 1 can irradiate the target user with X-rays to obtain an X-ray film containing the target user's pelvis and femur as femoral image data.

[0019] In some embodiments, the data processor 2 described above can be configured to perform the following processes: First, image detection processing is performed on the femoral image data received from the X-ray machine 1 to obtain the image detection result.

[0020] In some embodiments, the data processor 2 can perform image detection processing on the femoral image data received from the X-ray machine 1 to obtain an image detection result. The image detection result can be a label characterizing whether the position of the femoral image data is in the orthostatic position of the femur. For example, the image detection result can be "is in the orthostatic position of the femur" or "is not in the orthostatic position of the femur".

[0021] In some optional implementations of certain embodiments, the data processor 2 described above can be configured to perform the following steps: The first step is to preprocess the aforementioned femoral image data to obtain preprocessed femoral image data. This preprocessed femoral image data can be any pre-processed femoral image data. In practice, the data processor 2 first uses an image size adjustment function to adjust the image size of the femoral image data to a preset image size, obtaining the adjusted femoral image data as the adjusted femoral image data. This image size adjustment function can be any function capable of adjusting the image size. For example, the image size adjustment function can be the Resize() function in OpenCV. The preset image size can be a pre-defined image size. For example, the preset image size can be 224*224.

[0022] Secondly, the number of image channels in the femoral image data can be adjusted to a preset number using a channel adjustment function, resulting in adjusted femoral image data as preprocessed femoral image data. The channel adjustment function can be any function capable of adjusting the number of image channels. For example, the channel adjustment function could be `cv2.cvtColor()`. The preset number of channels can be a pre-defined value. For example, the preset number of channels could be 3.

[0023] The second step involves performing image classification processing on the preprocessed femoral image data to obtain image probability distribution data. This image probability distribution data characterizes the probability that the preprocessed femoral image data belongs to each category. The image probability distribution data can include image category labels. Each image category label corresponds to a confidence level. Each image category label can be a label representing the category of the body position corresponding to the preprocessed femoral image data. For example, the image category label can be "anteroposterior femoral view," "lateral femoral view," or "non-standard body position." The confidence level represents the probability that the preprocessed femoral image data belongs to the corresponding image category label.

[0024] As an example, the above image probability distribution data can be "{Image category label: femur anteroposterior view, confidence level: 0.9; Image category label: femur lateral view, confidence level: 0.05; Image category label: non-standard position, confidence level: 0.05}".

[0025] In practice, the data processor 2 described above can input the preprocessed femoral image data into a pre-trained image classification model to obtain image probability distribution data. The image classification model can be a neural network model that takes the preprocessed femoral image data as input and outputs the image probability distribution data. For example, the image classification model can be a preprocessed ResNet-50 (Residual Network 50). The preprocessing can be a process of fine-tuning the ResNet-50 model using the labeled preprocessed femoral image dataset and the cross-entropy loss function.

[0026] The third step is to determine the image detection result based on the above image probability distribution data.

[0027] In practice, firstly, the data processor 2 can determine the image category label with the highest confidence among the various image category labels included in the image probability distribution data as the target image category label. Secondly, in response to determining that the target image category label is "anteroposterior femur," "is anteroposterior femur" can be determined as the image detection result. Then, in response to determining that the target image category label is not "anteroposterior femur," "is not anteroposterior femur" can be determined as the image detection result.

[0028] Second, in response to determining that the above image detection results meet the preset image detection conditions, grayscale mapping processing is performed on the above femoral image data based on the preset historical femoral image dataset to obtain femoral grayscale image data.

[0029] In some embodiments, the data processor 2, in response to determining that the image detection result meets preset image detection conditions, performs grayscale mapping processing on the femoral image data based on a preset historical femoral image dataset to obtain femoral grayscale image data. The image detection conditions can be that the image detection result is "femoral in anteroposterior position". Each historical femoral image in the historical femoral image dataset can be a pre-acquired femoral image that meets the image detection conditions. The femoral grayscale image data can be femoral image data that has undergone grayscale processing.

[0030] In addressing the aforementioned technical issues in the application scenario of the intensive care unit (ICU), the following technical problems often arise: when using large models to output body position information, the system is prone to errors due to low image contrast and unclear details, leading to interruptions in subsequent body position adjustment processes. This necessitates system restarts or manual intervention before resuming operation, resulting in low system efficiency. Considering the specific requirements of this application scenario—namely, the need for rapid adjustment of user positions based on real-time status in the ICU, which demands high system efficiency—we have decided to adopt the following solution: In some optional implementations of certain embodiments, the data processor 2 described above can be configured to perform the following steps: The first step involves performing region recognition processing on each historical femoral image in the aforementioned historical femoral image dataset to obtain an intramedullary nail region matrix. This matrix can be a binary matrix of the same size as the historical femoral image dataset, where each element represents the category of a corresponding pixel in the historical femoral image. The intramedullary nail region matrix can consist of 0s and 1s. A "1" in the matrix represents a pixel whose category is intramedullary nail, and a "0" represents a pixel whose category is background.

[0031] In practice, for each historical femoral image in the aforementioned historical femoral image dataset, firstly, the data processor 2 can input the historical femoral image data into a pre-trained image segmentation model to obtain an intramedullary nail segmentation mask. This intramedullary nail segmentation mask can be a matrix of the same size as the historical femoral image data, where each element represents the probability that the corresponding pixel in the historical femoral image data belongs to the intramedullary nail category. The image segmentation model can be a neural network model that takes the historical femoral image data as input and the intramedullary nail segmentation mask as output. For example, the image segmentation model can be a preprocessed U-Net model. The preprocessing can be a process of fine-tuning the U-Net model using the labeled historical femoral image dataset and the cross-entropy loss function. Then, for each element value in the intramedullary nail segmentation mask, in response to determining that the element value is greater than a preset segmentation threshold, the element value can be replaced with a first preset value using a replacement function to update the intramedullary nail segmentation mask. This segmentation threshold can be a pre-set value between 0 and 1. The specific setting of the segmentation threshold is not limited here. The first preset value can be 1. The replacement function can be a function that replaces one piece of data with another. For example, the replacement function can be the replace() function. Then, in response to determining that the element value is less than or equal to the segmentation threshold, the element value can be replaced with a second preset value to update the intramedullary nail segmentation mask. The second preset value can be 0. Finally, the updated intramedullary nail segmentation mask can be determined as an intramedullary nail region matrix.

[0032] The second step involves generating brightness data for each intramedullary nail based on the obtained intramedullary nail region matrices and the aforementioned historical femoral image dataset. Each intramedullary nail brightness data point represents the average pixel value of the corresponding image region containing the intramedullary nail in the historical femoral image data.

[0033] In practice, for each of the aforementioned intramedullary nail region matrices, firstly, the data processor 2 can determine the historical femoral image data corresponding to the aforementioned intramedullary nail region matrix as the target historical femoral image data. Secondly, for each element value included in the aforementioned intramedullary nail region matrix, in response to determining that the aforementioned element value satisfies a preset region element condition, the pixel value whose position is the same as the position of the aforementioned element value among the pixel values ​​included in the target historical femoral image data can be determined as the target pixel value. For example, when the aforementioned element value is the element value in the third row and fifth column of the intramedullary nail region matrix, the target pixel value is the pixel value in the third row and fifth column of the target historical femoral image data. Here, the aforementioned region element condition can be an element value of 1. Then, the average value of the determined target pixel values ​​can be determined as the intramedullary nail brightness data corresponding to the aforementioned intramedullary nail region matrix. Thus, the brightness data of each intramedullary nail can be determined.

[0034] The third step is to generate the average brightness pixel value based on the brightness data of each intramedullary nail mentioned above. This average brightness pixel value can be the average of the brightness data of each intramedullary nail. In practice, the data processor 2 can determine the average of the brightness data of each intramedullary nail as the average brightness pixel value.

[0035] The fourth step is to determine the average value of each pixel value included in the above femoral image data as the pixel average value.

[0036] The fifth step is to determine the difference between the average brightness pixel value and the average pixel value as the pixel difference value.

[0037] Step 6: For each pixel value in the above femoral image data, the sum of the pixel value and the above pixel difference is determined as the pixel value, so as to update the femoral image data.

[0038] Step 7: Identify the updated femoral image data as femoral adjustment image data.

[0039] Step 8: Based on the aforementioned femoral adjustment image data, determine the minimum and maximum pixel values. The minimum pixel value can be the smallest pixel value in the aforementioned femoral adjustment image data. The maximum pixel value can be the largest pixel value in the aforementioned femoral adjustment image data. In practice, firstly, the data processor 2 can determine the largest pixel value among all pixel values ​​included in the aforementioned femoral adjustment image data as the maximum pixel value. Secondly, it can determine the smallest pixel value among all pixel values ​​included in the aforementioned femoral adjustment image data as the minimum pixel value.

[0040] Step 9: Determine the difference between the maximum pixel value and the minimum pixel value as the pixel interval value.

[0041] Step 10: Determine the pixel ratio as the ratio of the preset maximum value to the aforementioned pixel interval value. The preset maximum value can be a pre-set, maximum pixel value. For example, the preset maximum value can be 255.

[0042] Step 11: Based on the above pixel interval value and the above pixel ratio value, perform pixel value mapping processing on the above femoral adjustment image data to obtain the mapped femoral adjustment image data as femoral grayscale image data.

[0043] In practice, for each pixel value in the femoral adjustment image data, firstly, the data processor 2 can determine the difference between the pixel value and the minimum pixel value as the minimum pixel difference. Secondly, the product of the minimum pixel difference and the pixel ratio can be determined as the pixel value to update the femoral adjustment image data. Finally, the updated femoral adjustment image data can be determined as femoral grayscale image data.

[0044] The above-mentioned technical solution and related content, combined with the steps executed by the intelligent bed controller 3, serve as an inventive point of this disclosure, solving the problem of "low system operating efficiency." Factors leading to low system operating efficiency often include: when the system uses a large model to output body position information, it is prone to errors due to low image contrast and unclear details, leading to interruptions in subsequent body position adjustment processes. This necessitates a system restart or manual intervention before continued use, resulting in low system operating efficiency. Solving these factors can improve system operating efficiency. To achieve this effect, this disclosure firstly performs region recognition processing on each historical femoral image in the aforementioned historical femoral image dataset to obtain an intramedullary nail region matrix. This allows the identification of the image region where the intramedullary nail is located in the historical femoral image data. Secondly, based on the obtained intramedullary nail region matrices and the aforementioned historical femoral image dataset, brightness data for each intramedullary nail is generated. This allows the brightness of the image region where the intramedullary nail is located to be obtained. Then, based on the aforementioned intramedullary nail brightness data, a brightness pixel average is generated. Therefore, the average brightness of the image region where the intramedullary nail is located in each historical femoral image data can be obtained. Then, the average value of each pixel value included in the aforementioned femoral image data is determined as the pixel average. Next, the difference between the average brightness pixel value and the average pixel value is determined as the pixel difference. Then, for each pixel value included in the aforementioned femoral image data, the sum of the pixel value and the pixel difference is determined as the pixel value to update the femoral image data. This allows for the updating of the femoral image data. Then, the updated femoral image data is determined as femoral adjustment image data. Next, based on the aforementioned femoral adjustment image data, the minimum and maximum pixel values ​​are determined. Then, the difference between the maximum and minimum pixel values ​​is determined as the pixel interval value. Then, the ratio of the preset maximum value to the pixel interval value is determined as the pixel ratio. Finally, based on the pixel interval value and the pixel ratio, pixel value mapping processing is performed on the aforementioned femoral adjustment image data to obtain the mapped femoral adjustment image data as femoral grayscale image data. Therefore, the femoral adjustment image data can be adjusted to obtain femoral grayscale image data. Furthermore, because the image contrast and details of the femoral image data can be adjusted based on previously successfully recognized historical femoral image data before recognition, instead of directly using a large model to recognize the image and output body position information, the probability of system errors caused by low image contrast and unclear details can be reduced. This can decrease the number of system interruptions and thus improve system efficiency.

[0045] Third, the above-mentioned femoral grayscale image data is subjected to image standardization processing to obtain standardized femoral image data.

[0046] In some embodiments, the data processor 2 can perform image standardization processing on the femoral grayscale image data to obtain standardized femoral image data. The standardized femoral image data can be femoral grayscale image data that has undergone normalization and image channel expansion processing.

[0047] In practice, firstly, the data processor 2 can normalize the pixel values ​​in the femoral grayscale image data to between 0 and 1 using a normalization algorithm, obtaining normalized femoral grayscale image data as normalized femoral image data. The normalization algorithm can be any algorithm capable of normalizing images. For example, it could be Min-Max normalization. Then, the number of channels in the normalized femoral image data can be expanded to a preset number of image channels using a channel expansion function, obtaining expanded normalized femoral image data as standardized femoral image data. The channel expansion function can be any function capable of expanding the number of image channels. For example, it could be the `np.tile()` function. The image channel value can be 3.

[0048] As an example, when the normalized femoral image data is named "image", it can be copied 3 times in the dimension of color channels by "np.tile(image[:,:,None],(1,1,3))" to expand the channels of the normalized femoral image data and obtain the expanded normalized femoral image data as the standardized femoral image data.

[0049] Fourth, the standardized femoral image data is segmented to obtain individual femoral mask data.

[0050] In some embodiments, the data processor 2 can perform instance segmentation processing on the standardized femoral image data to obtain individual femoral mask data. Each femoral mask data can be data obtained after instance segmentation of the standardized femoral image data.

[0051] In addressing the technical problems mentioned above, the application scenario of in-home follow-up examinations for users with mobility impairments often presents the following challenges: wasted computational resources. Femoral image data contains numerous redundant areas. When the system processes this data using a large end-to-end model, each pixel in the image is processed at the same depth, leading to wasted computational resources when processing these redundant areas. Considering the specific requirements of this application scenario—the need for portable devices for in-home follow-up examinations—and the limited computing power of these devices (which rely solely on battery power), it is crucial to minimize wasted computational resources and extend battery life to complete the follow-up examination. Therefore, we have decided to adopt the following solution: In some optional implementations of certain embodiments, the data processor 2 described above can be configured to perform the following steps: The first step involves performing multi-scale feature extraction on the standardized femoral image data to obtain low-level, mid-level, and high-level semantic feature maps. The low-level, mid-level, and high-level semantic feature maps correspond to different image sizes and feature channel numbers.

[0052] In practice, the data processor 2 can input the standardized femoral image data into the multi-scale feature extraction network of a pre-trained image instance segmentation model to obtain low-level semantic feature maps, mid-level semantic feature maps, and high-level semantic feature maps. The image instance segmentation model can be a neural network model that takes the standardized femoral image data as input and outputs the corresponding femoral mask data.

[0053] The image instance segmentation model described above can include four layers.

[0054] The first layer can be a multi-scale feature extraction network. This multi-scale feature extraction network can be a neural network that takes standardized femoral image data as input and outputs low-level, mid-level, and high-level semantic feature maps corresponding to the standardized femoral image data. For example, the multi-scale feature extraction network mentioned above could be CSPDarkNet.

[0055] The aforementioned low-level semantic feature map can be the feature map with the smallest corresponding image size output by the aforementioned multi-scale feature extraction network.

[0056] The intermediate semantic feature map mentioned above can be the feature map output by the multi-scale feature extraction network, which has an image size larger than the low-level semantic feature map but smaller than the high-level semantic feature map.

[0057] The aforementioned high-level semantic feature map can be the feature map with the largest image size output by the aforementioned multi-scale feature extraction network.

[0058] The second layer can be a feature processing layer. This feature processing layer can include a feature enhancement network and a mask generation network.

[0059] The aforementioned feature enhancement network can be a neural network that takes low-level, mid-level, and high-level semantic feature maps as input and outputs low-level, mid-level, and high-level enhanced feature maps. For example, the aforementioned feature enhancement network can be a feature pyramid network. The aforementioned low-level enhanced feature map can be a low-level semantic feature map that fuses mid-level and high-level semantic feature maps. The aforementioned mid-level enhanced feature map can be a mid-level semantic feature map that fuses low-level and high-level semantic feature maps. The aforementioned high-level enhanced feature map can be a high-level semantic feature map that fuses low-level and mid-level enhanced feature maps. The aforementioned low-level, mid-level, and high-level enhanced feature maps have different image sizes but the same number of feature channels. The number of feature channels can be the number of channels in the feature map. The image size of the aforementioned low-level enhanced feature map is larger than the image size of the aforementioned mid-level enhanced feature map. The image size of the aforementioned mid-level enhanced feature map is larger than the image size of the aforementioned high-level enhanced feature map.

[0060] The aforementioned mask generation network can be a neural network that takes a high-level enhanced feature map as input and a prototype mask tensor as output. This mask generation network may include sampling convolutional layers, convolutional layers, and a normalization function.

[0061] The aforementioned sampling convolutional layer can be a convolutional layer that takes a high-level augmented feature map as input and a sampled feature map as output. Here, the sampled feature map can be a high-level augmented feature map whose image size has been adjusted to an initial size. The initial size can be the image size of the standardized femoral image data input to the multi-scale feature extraction network.

[0062] The aforementioned convolutional layer can be a neural network layer that takes a sampled feature map as input and an adjusted feature map as output. The adjusted feature map can be a sampled feature map whose corresponding number of feature channels has been adjusted to a preset channel value. The preset channel value can be a pre-defined numerical value. Here, the specific setting of the preset channel value is not limited.

[0063] The normalization function described above can be an activation function that takes an adjusted feature map as input and a prototype mask tensor as output. For example, the normalization function could be the Sigmoid function. The prototype mask tensor can be the image matrix corresponding to the adjusted feature map where each pixel value is compressed to between 0 and 1. The shape of the prototype mask tensor is the same as the shape of the adjusted feature map. As an example, when the shape of the adjusted feature map is (10, 10, 15), the corresponding shape of the prototype mask tensor is also (10, 10, 15), indicating that the size of the prototype mask tensor is 10*10, the corresponding number of feature channels is 15, or that the prototype mask tensor includes 15 two-dimensional matrices of size 10×10.

[0064] The third layer can be a classification and detection layer. This classification and detection layer can include a detection head, a decoding layer, and a filtering algorithm.

[0065] The aforementioned detection head can be a convolutional layer that takes preset anchor boxes, low-level augmented feature maps, mid-level augmented feature maps, and high-level augmented feature maps as input, and outputs the coordinates of each target anchor box and the data of each candidate bounding box. Each anchor box can be an anchor box set during model training. Each anchor box has a corresponding height and width. Each target anchor box coordinate can be the coordinates of the center point of the target anchor box in the standardized femoral image data. The target anchor boxes can be the anchor boxes used by the detection head within the aforementioned anchor boxes.

[0066] Each candidate bounding box in the above candidate bounding box data can be the data corresponding to a candidate bounding box. These candidate bounding boxes can be bounding boxes generated when predicting the category of pixels in low-level, mid-level, or high-level augmented feature maps. Each candidate bounding box in the above candidate bounding box data corresponds to an anchor box. Each candidate bounding box in the above candidate bounding box data can include bounding box offset data, confidence score, probability data for each category, and a sequence of mask coefficient data.

[0067] The bounding box offset data described above can be used to characterize the offset of the candidate bounding box relative to the anchor box. This offset data can include horizontal coordinate offset, vertical coordinate offset, height offset, and width offset. The horizontal coordinate offset can characterize the offset of the candidate bounding box relative to the anchor box in the horizontal direction. The vertical coordinate offset can characterize the offset of the candidate bounding box relative to the anchor box in the vertical direction. The height offset can characterize the offset of the candidate bounding box relative to the anchor box in terms of height. The width offset can characterize the offset of the candidate bounding box relative to the anchor box in terms of width.

[0068] As an example, the bounding box offset data mentioned above can be (0.1, -0.3, 0.4, 0.2), where "0.1" can be the horizontal coordinate offset, which can represent that the candidate bounding box is offset to the right by 0.1 relative to the anchor box; "-0.3" can be the vertical coordinate offset, which can represent that the candidate bounding box is offset downward by 0.3 relative to the anchor box; "0.4" can be the height offset, which can represent that the height of the candidate bounding box has increased by 0.4 relative to the anchor box; and "0.2" can be the width offset, which can represent that the width of the candidate bounding box has increased by 0.2 relative to the anchor box.

[0069] The confidence level in the candidate bounding box data can be used to characterize the probability that the image region corresponding to the candidate bounding box contains the target object to be identified. Each category probability data in the above category probability data can be used to characterize the probability that the image region in the candidate bounding box belongs to the corresponding category information. The above category information can be a label used to characterize the category of the image. For example, the above category information can be "intramedullary nail" or "femoral medullary cavity". Each category probability data in the above category probability data corresponds to one category information. Each mask coefficient data in the above mask coefficient data sequence can be a weight value generated from the candidate bounding box data. The number of each mask coefficient data in the above mask coefficient data sequence is the same as the number of feature channels of the above prototype mask tensor. For example, when the number of feature channels of the prototype mask tensor is 15, the number of each mask coefficient data included in the mask coefficient data sequence is also 15.

[0070] The aforementioned decoding layer can be a linear layer that takes the coordinates of each target anchor point and each candidate bounding box as input, and outputs the decoded bounding box data corresponding to each candidate bounding box as output. Each of the aforementioned bounding box decoded data can be processed candidate bounding box data. Each of the aforementioned bounding box decoded data can include bounding box coordinate data, confidence score, decoded category information, and a mask coefficient vector. The aforementioned bounding box coordinate data can be the pixel coordinates of the top-left and bottom-right vertices of the candidate bounding box in the standardized femoral image data. Each of the aforementioned bounding box decoded data can include the x-coordinate, y-coordinate, and x-coordinate of the top-left and top-right vertices of the candidate bounding box. The confidence score in the bounding box decoded data can be the confidence score in the candidate bounding box data. The aforementioned decoded category information can be the category information corresponding to the target category probabilities included in the candidate bounding box data. The aforementioned target category probabilities can be the category probability data with the highest corresponding value.

[0071] The aforementioned mask coefficient vector can be a row vector obtained by combining the mask coefficient data from the corresponding mask coefficient data sequence. The number of each element value in the aforementioned mask coefficient vector is the same as the number of feature channels of the prototype mask tensor.

[0072] In practice, for each candidate bounding box in the aforementioned candidate bounding box data, firstly, the decoding layer can determine the category probability with the largest value among all category probability data included in the candidate bounding box data as the target category probability. Secondly, the category information corresponding to the target category probability can be determined as the decoded category information.

[0073] Then, the decoding layer can determine the horizontal coordinate offset included in the candidate bounding box data as the target horizontal coordinate offset. Next, it can determine the anchor boxes corresponding to the candidate bounding box data as the anchor boxes to be processed. Then, it can determine the coordinates of the target anchor boxes corresponding to the candidate bounding box data as the coordinates of the anchor boxes to be processed. Then, the product of the target horizontal coordinate offset and the width of the anchor boxes to be processed can be determined as the first width. Then, the sum of the first width and the horizontal coordinate values ​​included in the coordinates of the anchor boxes to be processed can be determined as the center horizontal coordinate value.

[0074] Then, the ordinate offset included in the candidate bounding box data can be determined as the target ordinate offset. The product of the target ordinate offset and the height of the anchor point box to be processed can be determined as the first height. Then, the sum of the first height and the ordinate values ​​included in the coordinates of the anchor point box to be processed can be determined as the center ordinate value.

[0075] Then, the height offset included in the candidate bounding box data can be determined as the target height offset. Next, the height exponent data can be determined as the height exponent data, where the preset base can be the natural constant e. For example, when the target height offset is 0.3, the height exponent data can be e to the power of 0.3. Then, the product of the height of the anchor point box to be processed and the height exponent data can be determined as the bounding box height. Finally, half of the bounding box height can be determined as the second height.

[0076] Then, the width offset included in the candidate bounding box data can be determined as the target width offset. Next, the power of the target width offset with the preset base can be determined as the width exponent data. Then, the product of the width of the anchor box to be processed and the width exponent data can be determined as the bounding box width. Finally, half of the bounding box width can be determined as the second width.

[0077] Then, the difference between the aforementioned center x-coordinate value and the aforementioned second width can be determined as the x-coordinate value of the upper left corner. The difference between the aforementioned center y-coordinate value and the aforementioned second height can be determined as the y-coordinate value of the upper left corner. The sum of the aforementioned center x-coordinate value and the aforementioned second width can be determined as the x-coordinate value of the lower right corner. The sum of the aforementioned center y-coordinate value and the aforementioned second height can be determined as the y-coordinate value of the lower right corner.

[0078] Then, the x-coordinate of the top left corner, the y-coordinate of the top left corner, the x-coordinate of the bottom right corner, and the y-coordinate of the bottom right corner can be combined to form the bounding box coordinate data.

[0079] Finally, the bounding box coordinate data, the decoding category information, and the confidence scores included in the candidate bounding box data can be combined to form bounding box decoding data.

[0080] The aforementioned filtering algorithm can be an algorithm that takes the decoded data of each bounding box as input and the data of each target bounding box as output. For example, the aforementioned filtering algorithm can be a non-maximum suppression algorithm. Each target bounding box data can be the filtered decoded bounding box data. Each target bounding box data can include bounding box coordinate data, confidence score, decoding category information, and mask coefficient vector.

[0081] The fourth layer can be an output layer. This output layer may include a linear combination layer, a mask normalization function, a processing layer, and a mask output function.

[0082] The aforementioned linear combination layer can be a linear layer that takes the prototype mask tensor and the mask coefficient vectors included in each target bounding box data as input, and outputs each initial mask data. In practice, for each target bounding box data, the linear combination layer first determines the mask coefficient vector corresponding to the target bounding box data as the target mask coefficient vector. For each element value in the target mask coefficient vector, firstly, the order of the element values ​​in the target mask coefficient vector is determined as the order value. Secondly, the two-dimensional matrix with the index of the order value in each two-dimensional matrix included in the prototype mask tensor is determined as the target matrix. The product of the element values ​​and the target matrix is ​​determined as the target mask. Finally, the sum of the determined target masks is determined as the initial mask data corresponding to the target bounding box data. Thus, each initial mask data can be generated. The initial mask data corresponds one-to-one with the target bounding box data.

[0083] The mask normalization function described above can be a sigmoid function that takes initial mask data as input and outputs probability mask data. Each probability mask data can be an initial mask data whose element values ​​are scaled to between 0 and 1. There is a one-to-one correspondence between the probability mask data and the initial mask data.

[0084] The aforementioned processing layer can be a neural network layer that takes the coordinate data of each bounding box (including each target bounding box) as input and outputs each region mask data. Each region mask data can be a matrix obtained by processing the probability mask data. Each region mask data corresponds to one probability mask data and one target bounding box data. There is a one-to-one correspondence between region mask data, probability mask data, and target bounding box data.

[0085] In practice, for each of the aforementioned probability mask data, firstly, the processing layer can determine the bounding box coordinates corresponding to the probability mask data as the target coordinate data. Secondly, the probability mask data can be upsampled to the initial size using a bilinear interpolation algorithm to obtain upsampled mask data. Thirdly, a cropping function can be used to crop a submatrix corresponding to the target coordinate data from the upsampled mask data to obtain region mask data. The cropping function can be any function capable of cropping a specified submatrix from a matrix. For example, the crop() function could be used. As an example, when the target coordinate data is (3,4,6,7), the submatrix corresponding to the target coordinate data in the upsampled mask data can be a matrix composed of the elements from the third to sixth rows and the fourth to seventh columns of the upsampled mask data.

[0086] The aforementioned mask output layer can be a linear layer that takes region mask data as input and outputs femoral mask data. Each femoral mask data can be a region mask data whose element values ​​are all modified to 1 or 0. There is a one-to-one correspondence between femoral mask data and region mask data.

[0087] In practice, for each region mask data in the aforementioned region mask data, and for each element value included in the aforementioned region mask data, in response to determining that the element value is greater than a preset element threshold, the aforementioned mask output layer can modify the element value to a first value using a data modification function to update the region mask data. The aforementioned element threshold can be a preset value between 0 and 1. Here, the specific setting of the aforementioned element threshold is not limited. The aforementioned first value can be 1. The aforementioned data modification function can be a function capable of modifying the element values ​​in the matrix. For example, the aforementioned data modification function can be the modify() function. In response to determining that the element value is less than or equal to the aforementioned element threshold, the element value can be modified to a second value using the aforementioned data modification function to update the region mask data. The aforementioned second value can be 0. Finally, when all element values ​​in the region mask data have been processed, the updated region mask data can be determined as femoral mask data. Furthermore, the decoded category information included in the target bounding box data corresponding to each femoral mask data can be determined as the target category information of the femoral mask data.

[0088] The second step involves performing feature enhancement processing on the aforementioned low-level, mid-level, and high-level semantic feature maps to obtain low-level enhanced feature maps, mid-level enhanced feature maps, and high-level enhanced feature maps. In practice, the data processor 2 can input the aforementioned low-level, mid-level, and high-level semantic feature maps into the feature enhancement network in the feature processing layer of the aforementioned image instance segmentation model to obtain low-level enhanced feature maps, mid-level enhanced feature maps, and high-level enhanced feature maps.

[0089] The third step involves generating a prototype mask tensor based on the aforementioned high-level enhanced feature maps. In practice, the data processor 2 can input the high-level enhanced feature maps into the mask generation network in the feature processing layer of the image instance segmentation model to obtain the prototype mask tensor.

[0090] The fourth step involves performing classification and detection processing on the low-level, mid-level, and high-level enhanced feature maps based on the preset anchor boxes, obtaining the coordinates of each target anchor box and the data of each candidate bounding box. Each candidate bounding box includes bounding box offset data, confidence score, probability data for each class, and a sequence of mask coefficients. Each candidate bounding box corresponds to one anchor box. In practice, the data processor 2 can input the anchor boxes, low-level, mid-level, and high-level enhanced feature maps into the detection head of the classification and detection layer of the image instance segmentation model to obtain the coordinates of each target anchor box and the data of each candidate bounding box.

[0091] Fifth, based on the coordinates of each target anchor box, the candidate bounding box data is subjected to coordinate decoding processing to obtain the decoded bounding box data. Each bounding box decoded data includes bounding box coordinates, confidence score, decoded category information, and a mask coefficient vector. In practice, the data processor 2 can input the target anchor box coordinates and candidate bounding box data into the decoding layer of the classification and detection layer of the image instance segmentation model to obtain the decoded bounding box data.

[0092] Step 6: Based on the bounding box coordinate data included in the above-mentioned bounding box decoding data, non-maximum suppression processing is performed on the above-mentioned bounding box decoding data to obtain the target bounding box data. Each target bounding box data includes bounding box coordinate data, confidence score, decoding category information, and mask coefficient vector. In practice, the data processor 2 can input the above-mentioned bounding box decoding data into the filtering algorithm in the classification and detection layer of the above-mentioned image instance segmentation model to obtain the target bounding box data.

[0093] Step 7: Based on the aforementioned prototype mask tensor and the mask coefficient vectors included in the target bounding box data, generate initial mask data corresponding to each of the target bounding box data. In practice, the data processor 2 can input the aforementioned prototype mask tensor and the mask coefficient vectors included in the target bounding box data into the linear combination layer in the output layer of the image instance segmentation model to obtain the initial mask data.

[0094] Step 8: Normalize the initial mask data to obtain the probability mask data. In practice, the data processor 2 can input the initial mask data into the mask normalization function in the output layer of the image instance segmentation model to obtain the probability mask data.

[0095] Step 9: For each probability mask data in the above probability mask data, generate region mask data based on the probability mask data and the bounding box coordinate data corresponding to the probability mask data. In practice, the data processor 2 can input the above probability mask data and the bounding box coordinate data included in the above target bounding box data into the processing layer in the output layer of the above image instance segmentation model to obtain the region mask data.

[0096] Step 10: Binarize the generated region mask data to obtain individual femoral mask data. Each femoral mask data corresponds to a category. In practice, the data processor 2 can input the region mask data into the mask output layer of the image instance segmentation model to obtain individual femoral mask data.

[0097] The above-described technical solution and its related content, combined with the steps performed by the intelligent bed controller 3, serve as an inventive point of this disclosure, solving the problem of "waste of system computing resources." Factors leading to wasted system computing resources often include: the presence of numerous redundant regions in the femoral image data; and when the system processes the femoral image data using a large end-to-end model, the model processes each pixel in the image at the same depth, resulting in wasted computing resources when utilizing computing resources to process redundant regions. Solving these factors can reduce the waste of system computing resources. To achieve this, this disclosure first performs multi-scale feature extraction processing on the standardized femoral image data to obtain low-level semantic feature maps, mid-level semantic feature maps, and high-level semantic feature maps. Thus, three different scale feature maps can be extracted from the standardized femoral image data. Secondly, feature enhancement processing is performed on the low-level, mid-level, and high-level semantic feature maps to obtain low-level enhanced feature maps, mid-level enhanced feature maps, and high-level enhanced feature maps. Thus, feature enhancement processing can be performed on the extracted feature maps. Then, based on the aforementioned high-level enhanced feature maps, a prototype mask tensor is generated. This yields the prototype mask tensor. Next, based on preset anchor boxes, classification and detection processing is performed on the aforementioned low-level, mid-level, and high-level enhanced feature maps to obtain the coordinates of each target anchor box and the data of each candidate bounding box. Each candidate bounding box includes bounding box offset data, confidence score, probability data for each class, and a sequence of mask coefficient data. Each candidate bounding box corresponds to one anchor box. This allows for preliminary identification of standardized femoral image data. Then, based on the coordinates of each target anchor box, coordinate decoding processing is performed on the candidate bounding box data to obtain the decoded bounding box data. Each decoded bounding box includes bounding box coordinate data, confidence score, decoding class information, and a mask coefficient vector. This allows for coordinate decoding of each target anchor box to obtain the decoded bounding box data. Then, based on the bounding box coordinate data included in the aforementioned bounding box decoding data, non-maximum suppression processing is performed on the aforementioned bounding box decoding data to obtain various target bounding box data. Each target bounding box data includes bounding box coordinate data, confidence level, decoding class information, and a mask coefficient vector. This allows for the filtering of the obtained bounding box decoding data. Next, based on the aforementioned prototype mask tensor and the mask coefficient vectors included in the aforementioned target bounding box data, initial mask data corresponding to the aforementioned target bounding box data is generated. This yields various initial mask data. Then, the aforementioned initial mask data is normalized to obtain various probability mask data.Thus, various probability mask data can be obtained. Then, for each probability mask data, a region mask data is generated based on the probability mask data and the corresponding bounding box coordinates. Finally, the generated region mask data is binarized to obtain various femoral mask data, where each femoral mask data corresponds to target category information. Because candidate bounding box data can be identified from standardized femoral image data, and then filtered, only the image regions corresponding to the filtered target bounding box data are processed, rather than processing all pixels at the same depth, thus reducing the waste of computational resources when processing redundant regions.

[0098] Fifth, perform key point detection processing on the above femoral grayscale image data to obtain the coordinate data of each key point. Each key point coordinate data corresponds to a key point name.

[0099] In some embodiments, the data processor 2 can perform keypoint detection processing on the femoral grayscale image data to obtain keypoint coordinate data. Each keypoint coordinate data can be the pixel coordinates corresponding to a keypoint in the femoral grayscale image data. Each keypoint coordinate data also has a corresponding keypoint name. The keypoint name can be the name of the keypoint coordinate data. The keypoint name can include, but is not limited to: the tip of the intramedullary nail head, the proximal end point of the femoral neck, the distal end point of the femoral neck, the proximal end point of the femoral shaft, and the distal end point of the femoral shaft. The superior end point of the medial cortex of the femoral shaft, the inferior end point of the medial cortex of the femoral shaft, the superior end point of the medial cortex of the femoral neck, and the inferior end point of the medial cortex of the femoral neck.

[0100] The tip of the intramedullary nail mentioned above can be the tip of the head section of the intramedullary nail. The head section of the intramedullary nail can be the shorter nail in the intramedullary nail. The head section of the intramedullary nail can be referenced... Figure 3The proximal endpoint of the femoral neck can be a key point in the lower part of the femoral neck. The distal endpoint of the femoral neck can be a key point in the upper part of the femoral neck. The proximal endpoint of the femoral shaft can be a key point in the upper part of the femoral shaft. The distal endpoint of the femoral shaft can be a key point in the lower part of the femoral shaft. The superior endpoint of the medial cortex of the femoral neck can be a key point on the medial cortex of the femoral neck. The inferior endpoint of the medial cortex of the femoral neck can be a key point on the medial cortex of the femoral neck, located below the superior endpoint of the medial cortex of the femoral neck. The superior endpoint of the medial cortex of the femoral shaft can be a key point on the medial cortex of the femoral shaft. The inferior endpoint of the medial cortex of the femoral shaft can be a key point on the medial cortex of the femoral shaft, located below the superior endpoint of the medial cortex of the femoral shaft. The specific locations of each key point can be found by referring to [reference needed]. Figure 2 .

[0101] In addressing the technical problems mentioned above, and considering the application scenario during the winter icy season, the following technical issues arise: low system adjustment efficiency. When the system directly processes femoral image data using an end-to-end model, the model is easily affected by image noise and other factors, resulting in low accuracy of the generated positional information. Consequently, when adjusting the smart bed based on this positional information, repeated adjustments are required to position the user appropriately, leading to low system adjustment efficiency. Furthermore, considering the following requirements of this application scenario: during the winter icy season, prolonged road icing leads to a surge in fracture cases, resulting in a sharp increase in the number of people requiring post-fracture examinations. Therefore, the system needs to adjust the position of more users within a limited timeframe, resulting in a short time window for adjusting the position of each user and high demands on system efficiency. We have decided to adopt the following solution: In some optional implementations of certain embodiments, the data processor 2 described above can be configured to perform the following steps: The first step is to perform feature map extraction processing on the above-mentioned femoral grayscale image data to obtain the femoral feature map. The above-mentioned femoral feature map can be the feature map corresponding to the femoral grayscale image data.

[0102] In practice, the aforementioned data processor 2 can input the aforementioned femoral grayscale image data into the feature map extraction network of the pre-trained key point detection model to obtain the femoral feature map.

[0103] The aforementioned key point detection model can be a neural network model that takes femoral grayscale image data as input and outputs data of each key point.

[0104] The key point detection model mentioned above can include three layers.

[0105] The first layer can be a feature map extraction network. This feature map extraction network can be a neural network that takes femoral grayscale image data as input and outputs femoral feature maps. For example, the feature map extraction network can be a convolutional neural network.

[0106] The second layer can be a heatmap output layer. This heatmap output layer can be a convolutional layer that takes the femoral feature map as input and outputs heatmaps of various keypoints. Each keypoint heatmap can be the same size as the femoral grayscale image data and is used to represent the location of keypoints in the femoral grayscale image data. Each keypoint heatmap corresponds to a keypoint name. The keypoint name can be the name corresponding to the keypoint. The pixel value in each keypoint heatmap represents the probability that the corresponding pixel belongs to the keypoint corresponding to the keypoint name.

[0107] The third layer can be an output function. This output function can be a maximum value output function that takes the heatmap of each keypoint as input and outputs the data of each keypoint. This maximum value output function can be a function used to output the index of the data with the largest value among all data. For example, the maximum value output function could be the `argmax()` function. Each keypoint data point can be the pixel coordinates corresponding to the largest pixel value in the keypoint heatmap. Each keypoint data point also has a corresponding keypoint name.

[0108] The second step involves convolutional processing of the aforementioned femoral feature map to obtain heatmaps of each key point. In practice, the data processor 2 can input the aforementioned femoral feature map into the heatmap output layer of the aforementioned key point detection model to obtain heatmaps of each key point.

[0109] The third step involves generating keypoint data based on the aforementioned heatmaps of key points. Each keypoint data point corresponds to a keypoint name. In practice, the data processor 2 can input the aforementioned heatmaps of key points into the output function of the keypoint detection model to obtain the keypoint data.

[0110] The fourth step involves constructing a keypoint adjacency matrix based on a pre-defined initial matrix, pre-defined structural connection data, and the keypoint names corresponding to the aforementioned keypoint data. The structural connection data can be a list representing whether the keypoint data are interconnected. This structural connection data can include various connection relationships. Each connection relationship can be a label representing whether two keypoint data are connected. For example, a connection relationship can be "mutually connected" or "not connected." Each connection relationship corresponds to two keypoint data.

[0111] As an example, the above structural connection data can be {connection relationship: interconnected, key point data: proximal end of femoral neck, distal end of femoral neck}.

[0112] The keypoint adjacency matrix described above can be a binary matrix where each row corresponds to one keypoint data point, each column corresponds to one keypoint data point, and each element is a value representing whether the keypoint data points in the corresponding row and column are connected. The keypoint adjacency matrix can also be a matrix composed of 0s and 1s. A "0" in the keypoint adjacency matrix indicates that the keypoint data points in the corresponding row and column are not connected. A "1" in the keypoint adjacency matrix indicates that the keypoint data points in the corresponding row and column are connected.

[0113] The preset initial matrix can be a matrix with the number of rows equal to the number of keypoints, the number of columns equal to the number of keypoints mentioned above, and each element having a value of 0. The number of keypoints mentioned above can be the number of data points for each of the aforementioned keypoints. In the above initial matrix, each row corresponds to one keypoint name, and each column corresponds to one keypoint name.

[0114] In practice, for each connection relationship included in the above-mentioned structural connection data, in response to determining that the connection relationship satisfies a preset connection condition, firstly, the data processor 2 can determine the two key point data corresponding to the connection relationship as two target key point data. The connection condition can be that the connection relationship is "mutually connected". Secondly, the two key point names corresponding to the two target key point data can be determined as the first key point name and the second key point name, respectively.

[0115] Then, the two element values ​​in the initial matrix corresponding to the first keypoint name and the second keypoint name mentioned above can be identified as the two element values ​​to be modified. Specifically, the element value in the initial matrix whose row corresponds to the first keypoint name and whose column corresponds to the second keypoint name can be identified as the first element value to be modified. The element value in the initial matrix whose row corresponds to the second keypoint name and whose column corresponds to the first keypoint name can be identified as the second element value to be modified.

[0116] Next, the values ​​of the two elements to be modified can be changed to preset values ​​using a data modification function to update the initial matrix. This data modification function can be any function capable of modifying the numerical values ​​in the matrix. For example, the `modify()` function could be used. The preset modification value can be 1. Finally, after all the connection relationships have been processed, the updated initial matrix can be determined as the keypoint adjacency matrix.

[0117] Fifth, based on the aforementioned heatmaps and keypoint data, generate feature matrices for each node. Each node feature matrix can be a submatrix of the image matrix corresponding to the keypoint heatmap. Each node feature matrix corresponds to one keypoint data point.

[0118] In practice, for each key point data in the aforementioned key point data, the data processor 2 can determine the corresponding key point heatmap as the target key point heatmap. Next, the image matrix corresponding to the target key point heatmap can be determined as the target heatmap matrix. Then, a submatrix of a preset size centered on the location of the key point data within the target heatmap matrix can be determined as the target submatrix. The preset size can be a pre-defined size, for example, 5*5. Then, the target submatrix can be flattened along the row direction into a row vector as the initial vector using a flattening function. The flattening function can be a function capable of flattening a matrix along rows or columns. For example, the flatten() function can be used. Then, the key point data can be added to the initial vector using a data addition function to update the initial vector and determine the updated initial vector as the node feature matrix. The data addition function can be a function capable of adding data to a vector. For example, the add() function can be used.

[0119] Step 6: Based on the aforementioned keypoint adjacency matrix and the aforementioned node feature matrices, perform feature convolution processing on the aforementioned node feature matrices to obtain the respective convolutional node feature matrices. Each of the aforementioned convolutional node feature matrices can be a node feature matrix that has undergone convolution processing and aggregated the features of its neighboring nodes.

[0120] In practice, firstly, the data processor 2 can input the keypoint adjacency matrix and the feature matrices of each node into a pre-defined convolutional model to obtain the feature matrices of each convolutional node. The convolutional model can be a neural network that takes the keypoint adjacency matrix and the feature matrices of each node as input and outputs the feature matrices of each convolutional node. For example, the convolutional model can be a pre-trained graph convolutional neural network. The pre-training can be a process of fine-tuning the graph convolutional neural network using a node training dataset and a cross-entropy loss function. The training data for each node in the node training dataset can be the data used to train the convolutional model. The training data for each node in the node training dataset can include the keypoint adjacency matrix and the feature matrices of each node.

[0121] Step 7: Perform data mapping processing on the feature matrices of each convolutional node to obtain coordinate offset data. Each coordinate offset corresponds to a keypoint name. Each coordinate offset can be a coordinate representing the offset of the keypoint data. For example, a coordinate offset of (1, -3) indicates that the x-coordinate of the keypoint data needs to be increased by 1 and the y-coordinate by -3.

[0122] In practice, the data processor 2 described above can input the feature matrices of each convolutional node into a pre-trained fully connected layer to map the feature matrices of each convolutional node into coordinate offset data. The pre-training process can be a fine-tuning of the fully connected layer using the feature matrix set of each convolutional node and the mean squared error loss function.

[0123] Step 8: Based on the aforementioned keypoint data and coordinate offset data, generate the coordinate data for each keypoint. Each keypoint coordinate can be obtained by adding the keypoint data to its corresponding coordinate offset data. In practice, for each keypoint data, firstly, the data processor 2 can determine the keypoint name corresponding to the keypoint data as the name to be processed. Secondly, the coordinate offset data corresponding to the keypoint name in the coordinate offset data that is the name to be processed can be determined as the target coordinate offset data. Then, the sum of the keypoint data and the target coordinate offset data can be determined as the keypoint coordinate data.

[0124] The above technical solution and its related content, combined with the steps performed by the intelligent bed controller 3, serve as an inventive point of this disclosure, solving the problem of "low system adjustment efficiency." Factors leading to low system adjustment efficiency often include: when the system directly processes femoral image data through an end-to-end model, the model is easily affected by factors such as image noise, resulting in low accuracy of the generated positional information. Consequently, when adjusting the intelligent bed based on the positional information, repeated adjustments are required to position the user appropriately, leading to low system adjustment efficiency. Solving these factors can improve the system's adjustment efficiency. To achieve this effect, this disclosure first performs feature map extraction processing on the aforementioned femoral grayscale image data to obtain a femoral feature map. Secondly, the aforementioned femoral feature map undergoes convolution processing to obtain heatmaps of various key points. Then, based on the aforementioned heatmaps of various key points, data for each key point is generated, wherein each key point in the aforementioned key point data corresponds to a key point name. Therefore, based on the heatmaps of each key point, preliminary data for each key point can be generated. Then, based on a preset initial matrix, preset structural connection data, and the names of each key point corresponding to the aforementioned key point data, a key point adjacency matrix is ​​constructed. This allows for the construction of a key point adjacency matrix based on the connection relationships between key points. Next, based on the aforementioned heatmaps and key point data, feature matrices for each node are generated. Then, based on the aforementioned key point adjacency matrices and node feature matrices, feature convolution processing is performed on these node feature matrices to obtain convolutional node feature matrices. Then, data mapping processing is performed on these convolutional node feature matrices to obtain coordinate offset data, where each coordinate offset corresponds to a key point name. Finally, based on the aforementioned key point data and coordinate offset data, the final key point coordinate data is generated. Thus, based on the coordinate offset data and key point data, the final key point coordinate data can be generated. Because the identification of each key point involves first generating preliminary data for each key point, then constructing a key point adjacency matrix based on the connection relationships between each key point, and finally adjusting the key point data through the constraints of the key point adjacency matrix to generate the final coordinate data of each key point, instead of directly processing the femoral image data through a model, the accuracy of the identified key points can be improved. This, in turn, can improve the accuracy of the positional information generated based on each key point, reduce the number of adjustments to the smart hospital bed, and improve the system's adjustment efficiency.

[0125] Sixth, based on the above femoral mask data, the above key point coordinate data, and the names of the key points corresponding to the above key point coordinate data, generate the femoral detection results.

[0126] In some embodiments, the data processor 2 can generate femoral detection results based on the femoral mask data, the key point coordinate data, and the key point names corresponding to the key point coordinate data.

[0127] In some optional implementations of certain embodiments, the data processor 2 described above can be configured to perform the following steps: The first step is to generate the intramedullary nail filling degree based on the aforementioned femoral mask data. This intramedullary nail filling degree can be defined as the proportion of the intramedullary nail within the medullary cavity of the femur.

[0128] In practice, firstly, the data processor 2 can determine at least one femoral mask data whose target category information is "intramedullary nail" from the aforementioned femoral mask data as at least one intramedullary nail mask data. Secondly, it can determine each element value among the element values ​​included in the aforementioned at least one intramedullary nail mask data that satisfies a preset intramedullary nail mask condition as an intramedullary nail element value. The aforementioned intramedullary nail mask condition can be an element value of 1. Then, the number of the aforementioned intramedullary nail element values ​​can be determined as the number of intramedullary nail elements. Next, at least one femoral mask data whose target category information is "femoral medullary cavity" from the aforementioned femoral mask data can be determined as at least one target femoral mask data. Then, each element value among the element values ​​included in the aforementioned at least one target femoral mask data that satisfies a preset femoral mask condition can be determined as an individual femoral element value. The aforementioned femoral mask condition can be an element value of 1. Finally, the number of the aforementioned femoral element values ​​can be determined as the number of femoral elements. Finally, the ratio of the number of intramedullary nail elements to the number of femoral elements can be determined as the intramedullary nail filling degree.

[0129] The second step involves determining the proximal femoral neck, distal femoral neck, proximal femoral shaft, and distal femoral shaft endpoints based on the aforementioned keypoint coordinate data and their corresponding keypoint names. Specifically, the proximal femoral neck endpoint can be the keypoint coordinate data with the corresponding keypoint name "Proximal Femoral Neck Endpoint," the distal femoral neck endpoint can be the keypoint coordinate data with the corresponding keypoint name "Distal Femoral Neck Endpoint," the proximal femoral shaft endpoint can be the keypoint coordinate data with the corresponding keypoint name "Proximal Femoral Shaft Endpoint," and the distal femoral shaft endpoint can be the keypoint coordinate data with the corresponding keypoint name "Distal Femoral Shaft Endpoint."

[0130] In practice, firstly, the data processor 2 can determine the keypoint coordinates named "proximal femoral neck endpoint" from the aforementioned keypoint coordinate data as the proximal femoral neck endpoint. Secondly, it can determine the keypoint coordinates named "distal femoral neck endpoint" from the aforementioned keypoint coordinate data as the distal femoral neck endpoint. Then, it can determine the keypoint coordinates named "proximal femoral shaft endpoint" from the aforementioned keypoint coordinate data as the proximal femoral shaft endpoint. Finally, it can determine the keypoint coordinates named "distal femoral shaft endpoint" from the aforementioned keypoint coordinate data as the distal femoral shaft endpoint.

[0131] The third step involves generating a femoral neck axis vector based on the proximal and distal ends of the femoral neck. This femoral neck axis vector can be a vector pointing from the proximal end of the femoral neck to the distal end. In practice, the data processor 2 can determine the femoral neck axis vector as the difference between the distal and proximal ends of the femoral neck.

[0132] Fourth, based on the proximal and distal ends of the femoral shaft, a femoral shaft axis vector is generated. This femoral shaft axis vector can be a vector pointing from the proximal end to the distal end of the femoral shaft. In practice, the data processor 2 can determine the femoral shaft axis vector as the difference between the distal and proximal ends of the femoral shaft.

[0133] Fifth, based on the aforementioned femoral neck axis vector and femoral shaft axis vector, generate the neck-shaft angle data. The neck-shaft angle data can be the angle between the femoral neck axis normal vector and the femoral shaft axis vector. The femoral neck axis normal vector can be a vector perpendicular to the femoral neck axis vector.

[0134] In practice, firstly, the data processor 2 can determine the first data in the femoral neck axis vector as the first vector value. It can then determine the second data in the femoral neck axis vector as the second vector value. Next, the negative of the second value can be determined as the vector negative. Then, the vector negative and the first vector value can be combined into a planar vector as the femoral neck axis normal vector. For example, when the femoral neck axis vector is (1,2), the vector negative is -2, the first vector value is 1, and the resulting femoral neck axis normal vector is (-2,1). Then, the angle between the femoral neck axis normal vector and the femoral shaft axis vector can be determined as the vector angle. Secondly, in response to determining that the vector angle is obtuse, the vector angle can be determined as the neck-shaft angle data. Then, in response to determining that the vector angle is acute, the supplementary angle of the vector angle can be determined as the neck-shaft angle data.

[0135] Step 6: Based on the coordinate data of each key point mentioned above and the corresponding key point names, determine the tip of the intramedullary nail. The tip of the intramedullary nail can be the coordinate data of the key point named "Intramedullary Nail Tip". In practice, the executing entity can determine the coordinate data of the key point named "Intramedullary Nail Tip" from the above key point coordinate data as the tip of the intramedullary nail.

[0136] Step 7: Perform semantic segmentation on the aforementioned femoral head grayscale image data to obtain femoral head contour mask data. This femoral head contour mask data can be a matrix of the same size as the aforementioned femoral head grayscale image data, where each element can be a category number representing a pixel in the femoral head grayscale image data. The category number can be a corresponding number for a category. For example, the category number can be "0", representing that the corresponding pixel in the femoral head grayscale image data is in the category of background, or "1", representing that the corresponding pixel in the femoral head grayscale image data is in the category of femoral head contour.

[0137] In practice, the data processor 2 can input the aforementioned femoral head grayscale image data into a pre-trained femoral head contour detection model to obtain femoral head contour mask data. The femoral head contour detection model can be a neural network model that takes femoral head grayscale image data as input and femoral head contour mask data as output. For example, the femoral head contour detection model can be a pre-trained U-Net model. The pre-training process can be a fine-tuning of the U-Net model using a femoral head grayscale image dataset and a cross-entropy loss function.

[0138] Step 8: Based on the aforementioned femoral head contour mask data, generate each femoral head contour point. Each of these femoral head contour points can be the pixel coordinates corresponding to a pixel value categorized as a femoral head contour in the aforementioned femoral grayscale image data.

[0139] In practice, firstly, the data processor 2 can determine each element value in the femoral head contour mask data that satisfies the preset femoral head contour conditions as each target element value. Here, the femoral head contour conditions can be defined as element values ​​representing the category of the corresponding pixel value as a femoral head contour. For example, when category number "1" represents a femoral head contour, the femoral head contour condition can be an element value of "1". Secondly, for each target element value, the number of rows corresponding to the target element value can be determined as the target row number. The number of columns corresponding to the target element value can be determined as the target column number. Then, the target row number and the target column number can be combined into a two-dimensional coordinate as a femoral head contour point. Thus, each femoral head contour point can be determined.

[0140] Step 9: Based on the above-mentioned proximal end point of the femoral neck, the above-mentioned distal end point of the femoral neck, the above-mentioned various femoral head contour points, and the above-mentioned intramedullary screw tip points, generate tip-to-apex distance data.

[0141] Step 10: Based on the coordinate data of each key point and the name of each key point corresponding to the coordinate data of each key point, determine the upper medial end point of the femoral shaft, the lower medial end point of the femoral shaft, the upper medial end point of the femoral neck, and the lower medial end point of the femoral neck.

[0142] Specifically, the superior medial endpoint of the femoral shaft can be the coordinate data of the keypoint named "Superior Medial Cortical Endpoint of the Femoral Shaft". The inferior medial endpoint of the femoral shaft can be the coordinate data of the keypoint named "Inferior Medial Cortical Endpoint of the Femoral Shaft". The superior medial endpoint of the femoral neck can be the coordinate data of the keypoint named "Superior Medial Cortical Endpoint of the Femoral Neck". The inferior medial endpoint of the femoral neck can be the coordinate data of the keypoint named "Inferior Medial Cortical Endpoint of the Femoral Neck".

[0143] In practice, firstly, the data processor 2 can determine the coordinate data of the key point named "superior end point of the medial cortex of the femoral shaft" in the coordinate data of each key point as the superior end point of the medial cortex of the femoral shaft.

[0144] Secondly, the key point coordinate data with the name "medial subcortical endpoint of the femoral shaft" can be identified as the medial lower endpoint of the femoral shaft.

[0145] Then, the coordinate data of the key point named "superior end point of medial cortex of femoral neck" in the above key point coordinate data can be determined as the superior end point of medial femoral neck.

[0146] Finally, the key point coordinate data with the name "medial subcortical endpoint of femoral neck" can be identified as the medial inferior endpoint of the femoral neck.

[0147] Step 11: Based on the aforementioned superior medial end point of the femoral shaft, inferior medial end point of the femoral shaft, superior medial end point of the femoral neck, and inferior medial end point of the femoral neck, generate the cortical support result.

[0148] The aforementioned cortical support results can serve as labels to characterize the alignment of the medial cortex of the femoral neck and shaft. For example, the cortical support results can be labeled as "positive support," "neutral support," or "negative support."

[0149] Step 12: Based on the above intramedullary nail filling degree, the above cortical support results, the above tip-to-apex distance data, and the above neck-shaft angle data, generate the femoral examination results.

[0150] The femoral examination results described above can be the names of the diseases most likely to occur in the target user after undergoing intertrochanteric fracture surgery. For example, the femoral examination results could be, but are not limited to, "traumatic osteoarthritis," "nonunion," or "refracture after internal fixation." The femoral examination results also correspond to a target occurrence probability. This target occurrence probability can be the probability value of the target user developing the target disease. The target disease can be the disease characterized by the femoral examination results described above.

[0151] In practice, firstly, in response to determining that the cortical support result is "positive support," a preset first support value can be determined as support data. This first support value can be a numerical value that characterizes the cortical support result as "positive support." For example, the first support value can be 0. Secondly, in response to determining that the cortical support result is "neutral support," a preset second support value can be determined as support data. This second support value can be a numerical value that characterizes the cortical support result as "neutral support." For example, the second support value can be 1. Then, in response to determining that the cortical support result is "negative support," a preset third support value can be determined as support data. This third support value can be a numerical value that characterizes the cortical support result as "negative support." For example, the third support value can be 2.

[0152] Then, the determined support data, the intramedullary nail filling, the tip-to-apex distance data, and the neck-shaft angle data can be combined into a row vector as the femoral detection vector.

[0153] Then, the aforementioned femoral detection vector can be input into a pre-trained femoral classification and detection model to obtain the femoral detection results. The aforementioned femoral classification and detection model can be a neural network model that takes femoral detection data as input and femoral detection results as output.

[0154] The above femoral classification and detection model can include two layers.

[0155] The first layer can be a classification network. This classification network can be a neural network that takes femoral bone detection data as input and femoral bone probability distribution data as output. For example, the classification network can be a multilayer perceptron. The femoral bone probability distribution data can characterize the probability that the patient corresponding to the femoral bone detection data will develop various diseases after surgery. The femoral bone probability distribution data can include various occurrence probabilities. Each occurrence probability corresponds to a disease name. For example, the femoral bone probability distribution data can be [0.95, 0.04, 0.01], where "0.95", "0.04", and "0.01" each correspond to a disease name. The disease name can be the name of a disease. For example, the disease name can be, but is not limited to, "traumatic osteoarthritis", "fracture nonunion", or "refracture after internal fixation".

[0156] The second layer can be an output function. This output function can be an `argmax` function that takes femoral probability distribution data as input and femoral detection results as output. The femoral detection results can be the disease name corresponding to the target occurrence probability. The target occurrence probability can be the highest probability of occurrence among the corresponding values ​​in the femoral probability distribution data.

[0157] In some optional implementations of certain embodiments, the data processor 2 described above can be configured to perform the following steps: The first step is to generate femoral head contour information based on the aforementioned femoral head contour points. This femoral head contour information can be a standard equation for fitting a circle to the femoral head. The fitted circle can be a circle obtained by fitting the aforementioned femoral head contour points. In practice, the data processor 2 can use a fitting algorithm to fit the aforementioned femoral head contour points into a circle to obtain the femoral head contour information. This fitting algorithm can be any algorithm capable of fitting each point into a circle. For example, the fitting algorithm can be the least squares method.

[0158] The second step involves generating femoral neck connection information based on the proximal and distal femoral neck endpoints. This femoral neck connection information can be an expression for the straight lines corresponding to the proximal and distal femoral neck endpoints. In practice, the data processor 2 can substitute the proximal and distal femoral neck endpoints into a two-point equation of the straight line to obtain the femoral neck connection information.

[0159] The third step is to determine the perpendicular bisector of the femoral neck based on the above information on the femoral neck connection.

[0160] The aforementioned perpendicular bisector information of the femoral neck can be expressed as the expression corresponding to the perpendicular line connecting the proximal and distal ends of the femoral neck. In practice, firstly, the data processor 2 can determine the midpoint coordinates of the proximal and distal ends of the femoral neck as the midpoint of the femoral neck. Secondly, the slope of the femoral neck connection information can be determined as the femoral neck connection slope. Then, the ratio of the preset slope value to the slope of the femoral neck connection can be determined as the target slope. The preset slope value can be -1. Then, using the target slope as the slope of a straight line, the target slope and the midpoint of the femoral neck can be substituted into the general equation of a straight line to obtain the perpendicular bisector information of the femoral neck.

[0161] The fourth step is to determine the intersection of the above-mentioned vertical line information of the femoral neck and the above-mentioned femoral head contour information as the stress point of the femoral head.

[0162] The fifth step is to determine the distance between the above-mentioned force-bearing point of the femoral head and the tip of the above-mentioned intramedullary nail as the tip-to-apex distance data.

[0163] In some optional implementations of certain embodiments, the data processor 2 described above can be configured to perform the following steps: The first step involves generating a medial cortical vector of the femoral shaft based on the aforementioned superior medial endpoint and inferior medial endpoint of the femoral shaft. This medial cortical vector can be a vector pointing from the superior medial endpoint to the inferior medial endpoint of the femoral shaft. In practice, the data processor 2 can determine the medial cortical vector of the femoral shaft as the difference between the inferior medial endpoint and the superior medial endpoint.

[0164] The second step involves generating a medial femoral neck cortical vector based on the aforementioned superior medial endpoint and inferior medial endpoint of the femoral neck. This medial femoral neck cortical vector can be a vector pointing from the inferior medial endpoint to the superior medial endpoint. In practice, the data processor 2 can determine the medial femoral neck cortical vector as the difference between the superior and inferior medial endpoints of the femoral neck.

[0165] The third step involves generating the medial cortical angle based on the aforementioned medial cortical vector of the femoral shaft and the aforementioned medial cortical vector of the femoral neck. This medial cortical angle can be the angle between the aforementioned medial cortical vector of the femoral shaft and the aforementioned medial cortical vector of the femoral neck.

[0166] In practice, the data processor 2 can determine the angle between the medial cortical vector of the femoral shaft and the medial cortical vector of the femoral neck as the medial cortical angle.

[0167] Fourth, in response to determining that the aforementioned medial cortical angle satisfies a preset neutral support condition, the preset neutral support result is determined as the cortical support result. The aforementioned neutral support condition can be that the supplementary angle of the aforementioned medial cortical angle is less than a preset angle. The aforementioned preset angle can be a pre-defined angle. For example, the aforementioned preset angle can be 5 degrees. The aforementioned neutral support result can be a label used to characterize the aforementioned cortical support result as neutral. For example, the aforementioned neutral support result can be "neutral support".

[0168] Fifth, in response to the determination that the aforementioned medial cortical angle does not satisfy the aforementioned neutral support condition, the following steps are performed: The first sub-step involves determining the abscissa data of the femoral shaft based on the aforementioned superior medial end point of the femoral shaft. This abscissa data can be the abscissa value of the superior medial end point of the femoral shaft. In practice, the data processor 2 can determine the abscissa value of the superior medial end point of the femoral shaft as the abscissa data. For example, when the superior medial end point of the femoral shaft is (3,5), the first data "3" can be the abscissa value of the superior medial end point of the femoral shaft, and the second data "5" can be the ordinate value of the superior medial end point of the femoral shaft.

[0169] The second sub-step involves determining the femoral neck abscissa data based on the aforementioned medial inferior endpoint of the femoral neck. This abscissa data can be the abscissa value of the aforementioned medial inferior endpoint of the femoral neck. In practice, the data processor 2 can determine the abscissa value of the aforementioned medial inferior endpoint of the femoral neck as the femoral neck abscissa data.

[0170] The third sub-step involves determining, in response to the determination that the femoral shaft abscissa data is less than the femoral neck abscissa data, that the preset positive support result is defined as the cortical support result. Here, the positive support result can be a label used to characterize the cortical support result as positive. For example, the positive support result can be "positive support".

[0171] The fourth sub-step involves determining that the femoral shaft abscissa data is greater than the femoral neck abscissa data, and then defining the preset negative support result as the cortical support result. The negative support result can be a label used to characterize the negative cortical support result. For example, the negative support result can be "negative support".

[0172] In some embodiments, the intelligent bed controller 3 can be configured to control the intelligent bed to perform a position adjustment task based on the femoral bone detection results. The intelligent bed can be a bed capable of adjustment according to instructions. For example, the intelligent bed can be an intelligent nursing bed.

[0173] In some optional implementations of certain embodiments, the aforementioned intelligent bed controller 3 can be configured as follows: The first step involves generating femoral position information based on preset positioning prompts and the aforementioned femoral detection results. The positioning prompts can be information used to guide the multimodal large-scale model to output a suitable position for the user based on the femoral detection results. For example, the positioning prompt could be "Please provide a suitable position for the user based on the femoral detection results." The multimodal large-scale model can be DeepSeek. The femoral position information can be the name of the position generated based on the femoral detection results. For example, the femoral position information can be, but is not limited to: supine position, semi-recumbent position, or head-down, feet-up position. In practice, the intelligent bed controller 3 can input the positioning prompts and the femoral detection results into the multimodal large-scale model to obtain the text output by the multimodal large-scale model as the femoral position information.

[0174] The second step involves determining position adjustment information based on the aforementioned femoral position information. This position adjustment information can be an instruction used to control the smart bed to adjust the user's position to the position represented by the femoral position information. In practice, the smart bed controller 3 can send the aforementioned femoral position information to a target terminal. This target terminal can be a terminal used by a technician. Furthermore, it can receive the position adjustment information sent by the target terminal.

[0175] The third step involves controlling the smart bed to perform the position adjustment task based on the aforementioned position adjustment information. In practice, the smart bed controller 3 can send the aforementioned position adjustment information to the smart bed to control it to adjust the user's position to the position represented by the aforementioned femoral position information.

[0176] The above embodiments of this disclosure have the following beneficial effects: the intelligent bed control system based on femoral detection in some embodiments of this disclosure can reduce the system response time. Specifically, the reason for the long system response time is that when the femoral image data is processed end-to-end directly through the model, the computational complexity of the model is high, which can easily lead to excessive computational load on the system when running the model, causing the system to freeze and thus resulting in a long system response time. Based on this, in the intelligent bed control system based on femoral detection in some embodiments of this disclosure, firstly, an X-ray machine is configured to acquire femoral image data. Thus, femoral image data can be obtained. Then, a data processor is configured to perform the following steps: perform image detection processing on the femoral image data received from the X-ray machine to obtain an image detection result. Thus, the image can be detected. Secondly, in response to determining that the above image detection result meets the preset image detection conditions, grayscale mapping processing is performed on the above femoral image data based on a preset historical femoral image dataset to obtain femoral grayscale image data. Thus, the image is processed only when the detected image meets the processing conditions to obtain femoral grayscale image data. Then, the aforementioned femoral grayscale image data is subjected to image standardization processing to obtain standardized femoral image data. Next, instance segmentation processing is performed on the standardized femoral image data to obtain individual femoral mask data. Then, keypoint detection processing is performed on the aforementioned femoral grayscale image data to obtain coordinate data for each keypoint, where each keypoint coordinate corresponds to a keypoint name. Then, keypoint recognition is performed on the femoral grayscale image data to obtain coordinate data for each keypoint. Then, based on the aforementioned femoral mask data, the aforementioned keypoint coordinate data, and the corresponding keypoint names, a femoral detection result is generated. Finally, the intelligent bed controller is configured to control the intelligent bed to perform a patient positioning adjustment task based on the aforementioned femoral detection result. Therefore, the intelligent hospital bed can be controlled based on the generated femoral detection results to adjust the user's position. Because it first generates femoral mask data and key point coordinate data corresponding to the image, then uses this data to generate femoral detection results, and finally controls the intelligent hospital bed to perform position adjustment tasks based on these results—instead of directly generating position information corresponding to femoral image data from a large end-to-end model—the computational load on the system can be reduced. This decreases the frequency of system crashes due to excessive computational load and reduces system response time.

[0177] Figure 4 It can be used to visualize images after segmenting intramedullary nails and femoral medullary cavities. Figure 4 In this context, "femoral medullary cavity" can be represented by the category information of each pixel within a black bounding box and a black image region. "Intramedullary nail" can be represented by the category information of each pixel within a white bounding box and a white graphic region. Figure 4 The "L" in the designation indicates that the femur into which the intramedullary nail was implanted is the femur of the user's left leg.

[0178] In some implementations, clients and servers can communicate using any currently known or future-developed network protocol such as HTTP (Hypertext Transfer Protocol) and can interconnect with digital data communication (e.g., communication networks) of any form or medium. Examples of communication networks include local area networks (“LANs”), wide area networks (“WANs”), the Internet (e.g., the Internet of Things), and peer-to-peer networks (e.g., ad hoc peer-to-peer networks), as well as any currently known or future-developed networks.

[0179] The flowcharts and block diagrams in the accompanying drawings illustrate the architecture, functionality, and operation of possible implementations of systems, methods, and computer program products according to various embodiments of this disclosure. In this regard, each block in a flowchart or block diagram may represent a module, segment, or portion of code containing one or more executable instructions for implementing a specified logical function. It should also be noted that in some alternative implementations, the functions indicated in the blocks may occur in a different order than those indicated in the drawings. For example, two consecutively indicated blocks may actually be executed substantially in parallel, and they may sometimes be executed in reverse order, depending on the functions involved. It should also be noted that each block in the block diagrams and / or flowcharts, and combinations of blocks in the block diagrams and / or flowcharts, can be implemented using a dedicated hardware-based system that performs the specified function or operation, or using a combination of dedicated hardware and computer instructions.

[0180] The functions described above in this document can be performed at least in part by one or more hardware logic components. For example, exemplary types of hardware logic components that can be used, without limitation, include: field-programmable gate arrays (FPGAs), application-specific integrated circuits (ASICs), application-specific standard products (ASSPs), system-on-a-chip (SoCs), complex programmable logic devices (CPLDs), and so on.

[0181] The above description is merely a selection of preferred embodiments of this disclosure and an explanation of the technical principles employed. Those skilled in the art should understand that the scope of the invention involved in the embodiments of this disclosure is not limited to technical solutions formed by specific combinations of the above-described technical features, but should also cover other technical solutions formed by arbitrary combinations of the above-described technical features or their equivalents without departing from the above-described inventive concept. For example, technical solutions formed by substituting the above-described features with (but not limited to) technical features with similar functions disclosed in the embodiments of this disclosure.

Claims

1. An intelligent hospital bed control system based on femoral bone detection, comprising: The X-ray machine was configured to acquire femoral image data; The data processor is configured to perform the following steps: Image detection processing is performed on the femoral image data received from the X-ray machine to obtain image detection results; In response to determining that the image detection result meets the preset image detection conditions, the femoral image data is subjected to grayscale mapping processing based on the preset historical femoral image dataset to obtain femoral grayscale image data; The grayscale image data of the femur is subjected to image standardization processing to obtain standardized femur image data; The standardized femoral image data is segmented to obtain individual femoral mask data; The femoral grayscale image data is processed by key point detection to obtain the coordinate data of each key point, wherein each key point coordinate data corresponds to a key point name. Based on the femoral mask data, the coordinate data of each key point, and the names of each key point corresponding to the coordinate data of each key point, a femoral detection result is generated. The intelligent hospital bed controller is configured to control the intelligent hospital bed to perform position adjustment tasks based on the femoral bone detection results.

2. The system according to claim 1, wherein, The data processor is configured to: The femoral image data is preprocessed to obtain preprocessed femoral image data; The preprocessed femoral image data is subjected to image classification processing to obtain image probability distribution data, wherein the image probability distribution data includes each image category label, and each image category label corresponds to a confidence level; Based on the image probability distribution data, the image detection result is determined.

3. The system according to claim 1, wherein, The data processor is configured to: Based on the aforementioned femoral mask data, the intramedullary nail filling degree is generated; Based on the coordinate data of each key point and the name of each key point corresponding to the coordinate data of each key point, the proximal end of the femoral neck, the distal end of the femoral neck, the proximal end of the femoral shaft, and the distal end of the femoral shaft are determined. A femoral neck axis vector is generated based on the proximal end point and the distal end point of the femoral neck. A femoral shaft axis vector is generated based on the proximal end point and the distal end point of the femoral shaft. Based on the femoral neck axis vector and the femoral shaft axis vector, generate neck-shaft angle data; Based on the coordinate data of each key point and the name of each key point corresponding to the coordinate data of each key point, the tip of the intramedullary nail head is determined. Semantic segmentation processing is performed on the grayscale image data of the femur to obtain femoral head contour mask data; Based on the femoral head contour mask data, generate each femoral head contour point; Based on the proximal end point of the femoral neck, the distal end point of the femoral neck, each femoral head contour point, and the tip point of the intramedullary screw, the tip-to-apex distance data is generated; Based on the coordinate data of each key point and the name of each key point corresponding to the coordinate data of each key point, the medial superior end point of the femoral shaft, the medial inferior end point of the femoral shaft, the medial superior end point of the femoral neck, and the medial inferior end point of the femoral neck are determined. Based on the superior medial end point of the femoral shaft, the inferior medial end point of the femoral shaft, the superior medial end point of the femoral neck, and the inferior medial end point of the femoral neck, a cortical support result is generated; Based on the intramedullary nail filling degree, the cortical support result, the apex distance data, and the neck-shaft angle data, the femoral examination result is generated.

4. The system according to claim 3, wherein, The data processor is configured to: Based on the aforementioned femoral head contour points, femoral head contour information is generated; Based on the proximal end point and the distal end point of the femoral neck, femoral neck connection information is generated; Based on the femoral neck connection information, the femoral neck perpendicular line information is determined; The intersection of the vertical line information of the femoral neck and the contour information of the femoral head is determined as the stress point of the femoral head. The distance between the point of force application on the femoral head and the tip of the intramedullary nail is defined as the tip-to-apex distance.

5. The system according to claim 3, wherein, The data processor is configured to: Based on the superior medial end point and the inferior medial end point of the femoral shaft, a medial cortical vector of the femoral shaft is generated; Based on the superior medial end point and the inferior medial end point of the femoral neck, a medial cortical vector of the femoral neck is generated; Based on the medial cortical vector of the femoral shaft and the medial cortical vector of the femoral neck, the medial cortical angle is generated; In response to determining that the medial cortical angle satisfies a preset neutral support condition, the preset neutral support result is determined as the cortical support result; In response to determining that the medial cortical angle does not satisfy the neutral support condition, the following steps are performed: Based on the upper medial end point of the femoral shaft, determine the abscissa data of the femoral shaft; Based on the lower medial end point of the femoral neck, determine the abscissa data of the femoral neck; In response to determining that the femoral shaft abscissa data is less than the femoral neck abscissa data, the preset positive support result is determined as the cortical support result; In response to determining that the femoral shaft abscissa data is greater than the femoral neck abscissa data, the preset negative support result is determined as the cortical support result.

6. The system according to claim 1, wherein, The intelligent hospital bed controller is configured to: Based on the preset body position prompts and the femoral detection results, femoral body position information is generated; Based on the femoral position information, position adjustment information is determined; Based on the aforementioned body position adjustment information, the intelligent hospital bed is controlled to perform body position adjustment tasks.

Images