Medical intelligent infusion management system based on visual recognition

Abstract

The present invention belongs to the field of visual medical technology, and in particular relates to a medical intelligent infusion management system based on visual recognition, which includes a monitoring module and a transmission discrimination module; the monitoring module is composed of a distributed monitoring unit and a transmission enhancement unit, and can synchronously capture images of multiple infusion bags and record user information. By analyzing the number of images, network status and infusion rate, and using a transmission correction model to optimize the transmission efficiency of the image set, accurate labeling of the remaining liquid amount and time of the infusion bag is achieved. Secondly, the system evaluates the risk probability of delayed infusion bag replacement based on the corrected image set combined with user data; the risk probability is fed back to the transmission enhancement unit to adjust the image enhancement order and mark the delay risk in real time, and finally the enhanced and labeled image is displayed on the PAD device, helping medical staff to understand the infusion status in a timely manner and reduce the medical risks caused by untimely replacement of infusion bags.

Description

Technical Field

[0001] The present invention belongs to the field of visual medical technology, and in particular relates to a medical intelligent infusion management system based on visual recognition. Background Art

[0002] Intravenous infusion refers to the slow injection of drugs directly into the patient's veins over a period of time to achieve the treatment of various diseases. In order to ensure the safety and effectiveness of the infusion process, medical staff need to determine the infusion drip rate based on the patient's condition, age and the properties of the drugs used, and promptly monitor the position of the liquid level, replace or stop the input of the drug solution; currently, most infusion monitoring uses weighing infusion monitoring or infrared monitoring technology, among which weighing monitoring is affected by factors such as the material of the infusion container, liquid shaking, and drug addition, and has poor measurement accuracy. The installation of weighing equipment requires the modification of the infusion device, which increases costs and may damage the original facilities; infrared monitoring technology has complex liquid flow in the infusion tube, and bubbles and impurities can easily interfere with the infrared signal, leading to misjudgment; existing technologies also use visual technology for monitoring, but in the actual image transmission process, affected by factors such as network fluctuations and the number of simultaneous transmissions, the image may experience packet loss, distortion, etc., resulting in The image quality is degraded. For example, the Chinese patent with authorization announcement number CN114588402B discloses an infusion process monitoring system and its monitoring method based on visual recognition. The algorithm processing is complex and the hardware performance requirements are high. When the network traffic pressure is high, data processing is prone to delay, which affects the timeliness of infusion process judgment. In addition, the image recognition link requires a large amount of training data and complex algorithm optimization. The Chinese patent with publication number CN114596521A discloses an intravenous infusion monitoring method and system based on visual measurement. When identifying the liquid level position and drip rate, it involves multiple complex algorithms and is prone to recognition errors due to interference from light, infusion bottle shape and material. At the same time, both technical solutions require adjustment of the operating procedures, making it difficult to achieve seamless upgrades. Medical staff may not be very accepting of them during promotion. Summary of the Invention

[0003] In response to the shortcomings of the existing technology, the present invention proposes a medical intelligent infusion management system based on visual recognition, which includes a monitoring module and a transmission discrimination module; the monitoring module is composed of a distributed monitoring unit and a transmission enhancement unit, which can synchronously capture images of multiple infusion bags and record user information. By analyzing the number of images, network status and infusion rate, and using the transmission correction model to optimize the transmission efficiency of the image set, the remaining liquid amount and time of the infusion bag can be accurately marked. Secondly, the system evaluates the risk probability of delayed infusion bag replacement based on the corrected image set combined with user data; the risk probability is fed back to the transmission enhancement unit to adjust the image enhancement order in real time and mark the delay risk, and finally the enhanced and marked image is displayed on the PAD device, helping medical staff to understand the infusion status in a timely manner and reduce the medical risks caused by untimely replacement of infusion bags.

[0004] To achieve the above object, the present invention provides the following technical solutions:

[0005] A medical intelligent infusion management system based on visual recognition includes: a monitoring module and a transmission discrimination module; the monitoring module includes a distributed monitoring unit and a transmission enhancement unit;

[0006] The distributed monitoring unit collects synchronous images, user information, and network transmission status of at least one or more infusion bags. Based on the number of synchronously collected images, the network transmission status, and the infusion rate, combined with the transmission correction model built into the transmission enhancement unit, a synchronous transmission-corrected infusion bag image set is obtained and the remaining liquid volume and remaining infusion time are marked.

[0007] Based on the corrected infusion bag image set and user information, the transmission discrimination module is used to obtain the infusion bag replacement delay risk probability;

[0008] The delay risk probability of the infusion bag replacement is fed back to the transmission enhancement unit, the transmission image enhancement sequence is adjusted in real time and the delay risk probability is marked, and the enhanced and marked image is displayed on the PAD in real time.

[0009] Specifically, the construction process of the transmission correction model includes:

[0010] According to the synchronously acquired infusion bag image, the infusion bag image is segmented by the background segmentation layer to obtain the infusion bag segmentation region set and the background segmentation region set, and the liquid level line coordinates in the infusion bag segmentation region are marked;

[0011] Based on the infusion bag shape attribute information, infusion bag segmentation area, full liquid volume and initial liquid level, a three-dimensional model of the infusion bag is obtained through the three-dimensional modeling layer;

[0012] Based on the three-dimensional model of the infusion bag and different infusion rates, the liquid surface separation line position function under different infusion rates was fitted by using a simulation algorithm combined with the support vector machine kernel function to obtain the separation line position function corresponding to different infusion rates.

[0013] Specifically, the construction process of the transmission correction model also includes:

[0014] Masking the liquid separation lines in the infusion bag segmentation area, inputting each masked infusion bag segmentation area and the liquid separation line position function into the feature extraction layer of the transmission correction model, and obtaining the feature space of each infusion bag segmentation area and the liquid level coordinate space predicted by the liquid separation line position function;

[0015] The liquid level line position prediction error loss is obtained based on the liquid level line coordinate space predicted by the infusion bag segmentation area feature space and the liquid separation line position function, combined with the marked liquid level line coordinates.

[0016] Specifically, the construction process of the transmission correction model also includes:

[0017] The continuous infusion bag image sequence constructed by masking the infusion bag segmentation area at consecutive time points is input into the optical flow layer. The motion vector of each pixel corresponding to the consecutive time points of the masked infusion bag segmentation area is analyzed to obtain the dense optical flow field corresponding to the continuous masked infusion bag segmentation area.

[0018] Based on the change values of adjacent pixels in the dense optical flow field corresponding to each masked infusion bag segmentation area, a clustering algorithm combined with the quarter-third quantile is used to construct the dynamic clustering sub-region pixel band space corresponding to each masked infusion bag segmentation area.

[0019] According to the pixel change value in the dynamic clustering sub-region pixel band space and the pixel change mean corresponding to the full-load liquid infusion bag, a pixel change abnormality threshold is constructed to obtain the abnormal dynamic clustering sub-region pixel band;

[0020] The abnormal dynamic clustering sub-region pixel band is a dynamic clustering sub-region pixel band corresponding to the largest number of abnormal pixel change coordinate points and the largest abnormal deviation value.

[0021] Specifically, the construction process of the transmission correction model also includes:

[0022] Input the pixel band of the abnormal dynamic clustering sub-region into the Gaussian kernel convolution, diffuse the corresponding abnormal pixel points in the pixel band of the abnormal dynamic clustering sub-region, and obtain the continuous abnormal pixel point coordinate space;

[0023] Feeding back the continuous abnormal pixel point coordinate space to the feature extraction layer to perform secondary correction on the liquid level line position prediction error loss and the predicted liquid level line coordinate space to obtain the corrected liquid level line position prediction error loss and the predicted liquid level line coordinate space;

[0024] The infusion bag segmentation area space to be enhanced is constructed based on the corrected liquid level line position prediction error loss, the predicted liquid level line coordinate space, and the infusion bag segmentation area feature space.

[0025] Specifically, the construction process of the transmission correction model also includes:

[0026] Based on the available bandwidth, packet loss rate, and transmission delay corresponding to the number of images transmitted simultaneously in different time periods in the network transmission status, the network quality evaluation layer is used to construct the network quality index corresponding to different time periods.

[0027] Based on the network quality index corresponding to different time periods and the preset network quality level interval, the network transmission level corresponding to different time periods is obtained;

[0028] The network transmission levels corresponding to the different time periods and the preset image enhancement resolution table are input into the hierarchical correction layer to simulate the enhanced transmission of the segmented area space of the infusion bag to be enhanced, so as to obtain the transmission delay and packet loss of each resolution image at different network transmission levels and the corresponding network level-image quality mapping matrix.

[0029] Specifically, the construction process of the transmission correction model also includes:

[0030] Based on the different pixel spaces of the masked infusion bag segmentation areas and the predicted liquid level position combined with the real-time infusion rate, a probability evaluation layer with a built-in Bayesian function is used to obtain the replacement delay risk probability value and probability evaluation loss corresponding to each infusion bag at the current moment.

[0031] The replacement delay risk probability value corresponding to each infusion bag at the current moment is used as the processing priority of the infusion bag segmentation area space to be enhanced corresponding to each infusion bag captured image;

[0032] The processing priority of the space of the infusion bag segmentation area to be enhanced corresponding to each infusion bag captured image is fed back to the simulated enhancement transmission process, and the optimal image quality corresponding to each processing priority image under different network levels is simulated. The transmission delay is used as a measurement indicator to construct a processing priority-network level-image quality-transmission delay mapping matrix space.

[0033] Specifically, the construction process of the transmission correction model also includes:

[0034] The processing priority-network level-image quality-transmission delay mapping matrix space is built into the output layer. The infusion bag segmentation area space to be enhanced, the network quality index, and the processing priority corresponding to each infusion bag at the current moment are input into the output layer to obtain the infusion bag image of the corresponding resolution image quality and the corresponding image enhancement loss;

[0035] The loss function of the transmission correction model is constructed based on the corrected liquid level position prediction error loss, probability assessment loss, transmission delay and packet loss loss, and image enhancement loss. The transmission correction model is simulated and trained, and the transmission delay corresponding to each resolution image quality under different network quality and processing priority is analyzed.

[0036] When the transmission delay corresponding to each resolution image quality is less than the preset transmission delay, the trained transmission correction model is obtained, and the enhanced infusion bag image and the corresponding enhanced resolution at the corresponding network level are output.

[0037] Specifically, the process of masking the liquid separation line in the segmentation area of the infusion bag includes:

[0038] Based on the segmented region of the infusion bag, a multi-level pyramid feature map corresponding to the segmented region of the infusion bag is generated by a Gaussian pyramid decomposition algorithm combined with a preset number of layers;

[0039] Based on the multi-level pyramid feature map, the horizontal gradient and vertical gradient convolution are performed on each level of the pyramid feature map through the Sobel operator convolution to obtain the horizontal and vertical gradient amplitudes and gradient direction angles corresponding to each level of the pyramid feature map;

[0040] Based on the gradient direction angle and the preset gradient direction angle filtering interval, the horizontal and vertical gradient amplitudes are filtered to obtain the filtered horizontal and vertical gradient amplitudes corresponding to each level of the pyramid feature map;

[0041] The horizontal and vertical gradient amplitudes corresponding to the filtered feature maps of each level of the pyramid are weighted attenuated through the channel convolution attention algorithm, and the feature maps of each level of the pyramid completed by weighted attenuation are upsampled and fused through bilinear interpolation to obtain a multi-level pyramid fusion gradient map.

[0042] Specifically, the process of masking the liquid separation line in the segmentation area of the infusion bag further includes:

[0043] Based on the multi-level pyramid fusion gradient image, the mean and variance of the fused gradient image are obtained through statistical algorithms;

[0044] The mask interval is set based on the mean and variance corresponding to the fused gradient map, and the initial binary mask corresponding to the liquid level line in the infusion bag segmentation area is generated according to the set mask interval combined with non-maximum suppression;

[0045] Based on the initial binary mask corresponding to the liquid level line in the infusion bag segmentation area, the binary mask corresponding to the liquid level line in the infusion bag segmentation area is obtained through the 3×3 rectangular structure element S combined with the connected domain filter. The liquid level line marked in the input infusion bag segmentation area is then binary masked based on the binary mask corresponding to the liquid level line in the liquid bag segmentation area.

[0046] Compared with the prior art, the present invention has the following beneficial effects:

[0047] The present invention addresses the deficiencies of the existing technology and solves the problems of image packet loss and distortion caused by factors such as network fluctuations and the number of synchronous transmissions in the existing visual recognition technology during infusion management by introducing a distributed monitoring unit and a transmission enhancement unit. Specifically, the present invention can efficiently and synchronously capture images and record user information for multiple infusion bags, and optimize the image transmission quality using a transmission correction model to ensure that a clear and accurate set of infusion bag images can be obtained even when the network status is poor, and accurate annotation of the remaining liquid volume and infusion time can be achieved. In addition, the transmission discrimination module is combined to evaluate and feedback the probability of infusion bag replacement delay risk, adjust the image enhancement sequence and marking delay risk in real time, and finally display the enhanced and annotated image on the PAD, effectively improving the work efficiency of medical staff and the treatment safety of patients. Therefore, compared with the existing technology, the present invention improves the stability and accuracy of image transmission and reduces medical risks. BRIEF DESCRIPTION OF THE DRAWINGS

[0048] Figure 1 This is a flow chart of the medical intelligent infusion management system based on visual recognition according to Example 1 of the present invention;

[0049] Figure 2 This is a diagram of the construction architecture of the transmission correction model of Example 1 of the present invention. DETAILED DESCRIPTION

[0050] Example 1

[0051] Intravenous infusion is a common and effective means of treating diseases. It involves injecting drugs directly into the patient's veins at a slow rate over a specific period of time. In the daily operation of hospitals, nursing resources are relatively limited, but they have to deal with a large number of infusion patients, and most patients do not have dedicated care during the infusion. In this case, how to accurately and efficiently transmit the real-time status of the infusion bag, such as the remaining amount of medicine, infusion rate and other information, clearly and quickly to the nurse side has become a key issue that needs to be solved urgently. If this information cannot be obtained in a timely manner, it will easily lead to delays in fluid replacement, causing blood backflow in the infusion tube, which will not only affect the treatment process, but may also cause discomfort and even additional health risks to the patient. At the same time, delays in removing the drip may also cause the patient to maintain the same posture for a long time, causing physical fatigue, and occupying limited medical resources; for this, please refer to Figure 1 , an embodiment provided by the present invention: a medical intelligent infusion management system based on visual recognition, comprising: a monitoring module and a transmission discrimination module; the monitoring module comprises a distributed monitoring unit and a transmission enhancement unit;

[0052] The monitoring module is used for synchronous acquisition of infusion bag images and transmission correction of acquired images; the distributed monitoring unit, according to the configured synchronous monitoring model, synchronously acquires at least one or more infusion bag images, user information, and network transmission status during image transmission in real time;

[0053] The transmission enhancement unit corrects the synchronously transmitted infusion bag image set and marks the remaining liquid volume and remaining infusion time according to the number of synchronously collected images, network transmission status, infusion rate and the configured transmission correction model;

[0054] The transmission enhancement unit obtains the risk probability of delayed infusion bag replacement based on the corrected infusion bag image set and user information through a Bayesian function, and feeds the risk probability of delayed infusion bag replacement back to the transmission enhancement unit, adjusts the transmission image enhancement sequence in real time and annotates the delay risk probability, and displays the enhanced and annotated image on a PAD (portable electronic device) in real time.

[0055] Furthermore, the workflow corresponding to the distribution monitoring unit, the transmission enhancement unit, and the transmission determination module in this embodiment includes:

[0056] The distributed monitoring unit collects synchronous images, user information, and network transmission status of at least one or more infusion bags. Based on the number of synchronously collected images, the network transmission status, and the infusion rate, combined with the transmission correction model built into the transmission enhancement unit, a synchronous transmission-corrected infusion bag image set is obtained and the remaining liquid volume and remaining infusion time are marked.

[0057] Based on the corrected infusion bag image set and user information, the transmission discrimination module is used to obtain the infusion bag replacement delay risk probability;

[0058] The delay risk probability of the infusion bag replacement is fed back to the transmission enhancement unit, the transmission image enhancement sequence is adjusted in real time and the delay risk probability is marked, and the enhanced and marked image is displayed on the PAD in real time.

[0059] Further, see Figure 2 The construction process of the transmission correction model in this embodiment includes:

[0060] According to the synchronously acquired infusion bag image, the infusion bag image is segmented by the background segmentation layer to obtain the infusion bag segmentation region set and the background segmentation region set, and the liquid level line coordinates in the infusion bag segmentation region are marked;

[0061] Furthermore, in this embodiment, the specific process of background segmentation and labeling includes:

[0062] The U-Net encoder is constructed using depthwise separable convolution, and the U-Net decoder is constructed using transposed convolution. The synchronously collected infusion bag image is input into the constructed U-Net encoder and gradually downsampled to the original image through the convolution layer with a step size of 2. Figure 1 / 32 size, and input the downsampled infusion bag image features into the U-Net decoder and splice them with the feature maps of the corresponding layer of the encoder. During splicing, SE attention is used to align the channels to obtain the decoded infusion bag image features;

[0063] The decoded infusion bag image features are combined with Dice Loss and Focal Loss, and the U-Net output sublayer is used to obtain the infusion bag segmentation region set. In the segmented infusion bag region, a semi-automatic labeling tool (such as LabelMe) is used to calibrate the liquid level line coordinates to generate a labeled dataset.

[0064] Furthermore, the process of constructing the three-dimensional model of the infusion bag in this embodiment includes:

[0065] Calculate the depth map of the infusion bag surface through stereo matching or laser speckle displacement, convert the depth map into a 3D point cloud, and optimize the deformation model by combining the shape properties of the infusion bag (such as material flexibility parameters);

[0066] Based on the infusion bag shape attribute information, infusion bag segmentation area, full liquid volume and initial liquid level, a three-dimensional model of the infusion bag is obtained through the three-dimensional modeling layer;

[0067] Based on the three-dimensional model of the infusion bag and different infusion rates, the liquid surface separation line position function under different infusion rates was fitted by using a simulation algorithm combined with the support vector machine kernel function to obtain the separation line position function corresponding to different infusion rates.

[0068] Furthermore, the liquid dividing line in this embodiment refers to the dividing line corresponding to the liquid level in the infusion bag.

[0069] Masking the liquid separation lines in the infusion bag segmentation area, inputting each masked infusion bag segmentation area and the liquid separation line position function into the feature extraction layer of the transmission correction model, and obtaining the feature space of each infusion bag segmentation area and the liquid level coordinate space predicted by the liquid separation line position function;

[0070] Furthermore, the process of masking the liquid separation line in the segmentation area of the infusion bag in this embodiment includes:

[0071] Based on the segmented region of the infusion bag, a multi-level pyramid feature map corresponding to the segmented region of the infusion bag is generated by a Gaussian pyramid decomposition algorithm combined with a preset number of layers;

[0072] Based on the multi-level pyramid feature map, the horizontal gradient and vertical gradient convolution are performed on each level of the pyramid feature map through the Sobel operator convolution to obtain the horizontal and vertical gradient amplitudes and gradient direction angles corresponding to each level of the pyramid feature map;

[0073] Based on the gradient direction angle and the preset gradient direction angle filtering interval, the horizontal and vertical gradient amplitudes are filtered to obtain the filtered horizontal and vertical gradient amplitudes corresponding to each level of the pyramid feature map;

[0074] The horizontal and vertical gradient amplitudes corresponding to the filtered feature maps of each level of the pyramid are weighted attenuated through the channel convolution attention algorithm, and the feature maps of each level of the pyramid completed by weighted attenuation are upsampled and fused through bilinear interpolation to obtain a multi-level pyramid fusion gradient map.

[0075] Furthermore, the process of obtaining the multi-level pyramid feature map in this embodiment includes:

[0076] The Gaussian pyramid decomposition algorithm is combined with the preset 3-layer decomposition number to perform step-by-step downsampling to generate a three-level pyramid feature map with σ=1, 2, and 3;

[0077] Apply the Sobel operator on each layer of the pyramid feature map to extract horizontal and vertical gradients and retain the edge response in the horizontal direction (±10°);

[0078] Furthermore, in this embodiment, σ is the standard deviation of the Gaussian kernel;

[0079] Furthermore, the step-by-step downsampling process in this embodiment includes:

[0080] Each layer of the pyramid image is obtained by Gaussian blurring the image of the previous layer and then downsampling (e.g., reducing the size by half);

[0081] For example:

[0082] Original image (blurred with σ=1) → downsampled to get layer 1.

[0083] Layer 1 image (blurred by σ=2) → downsampled to get layer 2.

[0084] Layer 2 image (blurred by σ=3) → downsampled to get layer 3.

[0085] The blur layer with σ=1:

[0086] Used for precise positioning of the liquid level, suppressing interference from small bubbles or slight reflections.

[0087] Fuzzy layer with σ=2:

[0088] Analyze the overall deformation of the infusion bag label area to avoid the influence of local noise.

[0089] Fuzzy layer with σ=3:

[0090] Extract the global outline of the infusion bag to assist in 3D modeling and liquid volume estimation.

[0091] The multi-scale gradient map is restored to its original size through bilinear interpolation to obtain a multi-level pyramid fusion gradient map. Furthermore, this embodiment superimposes each level of the pyramid feature map according to weights (0.6, 0.3, 0.1) to enhance the continuity of the liquid level line.

[0092] Based on the multi-level pyramid fusion gradient image, the mean and variance of the fused gradient image are obtained through statistical algorithms;

[0093] The mask interval is set based on the mean and variance corresponding to the fused gradient map, and the initial binary mask corresponding to the liquid level line in the infusion bag segmentation area is generated according to the set mask interval combined with non-maximum suppression;

[0094] Based on the initial binary mask corresponding to the liquid level line in the infusion bag segmentation area, the binary mask corresponding to the liquid level line in the infusion bag segmentation area is obtained through the 3×3 rectangular structure element S combined with the connected domain filter. The liquid level line marked in the input infusion bag segmentation area is then binary masked based on the binary mask corresponding to the liquid level line in the liquid bag segmentation area.

[0095] Furthermore, the detailed process of obtaining the binary mask in this embodiment includes:

[0096] Mean μ and variance based on fused gradient map s , set the high threshold (μ+1.5s) and low threshold (μ-0.5s) corresponding to the pixel filtering interval;

[0097] Based on the pixel filtering interval, pixels above the high threshold are retained as seed points, and the areas above the low threshold are connected to form the initial mask;

[0098] A 3×3 rectangular kernel is used to close the initial mask (dilation followed by erosion) to fill the gaps. Noise is filtered out through connected domain analysis (area > 20 pixels) to obtain a binary mask corresponding to the liquid level line in the segmented area of the infusion bag.

[0099] The liquid level line position prediction error loss is obtained based on the liquid level line coordinate space predicted by the infusion bag segmentation area feature space and the liquid separation line position function, combined with the marked liquid level line coordinates.

[0100] The continuous infusion bag image sequence constructed by masking the infusion bag segmentation area at consecutive time points is input into the optical flow layer. The motion vector of each pixel corresponding to the consecutive time points of the masked infusion bag segmentation area is analyzed to obtain the dense optical flow field corresponding to the continuous masked infusion bag segmentation area.

[0101] Based on the change values of adjacent pixels in the dense optical flow field corresponding to each masked infusion bag segmentation area, a clustering algorithm combined with the quarter-third quantile is used to construct the dynamic clustering sub-region pixel band space corresponding to each masked infusion bag segmentation area.

[0102] According to the pixel change value in the dynamic clustering sub-region pixel band space and the pixel change mean corresponding to the full-load liquid infusion bag, a pixel change abnormality threshold is constructed to obtain the abnormal dynamic clustering sub-region pixel band;

[0103] The abnormal dynamic clustering sub-region pixel band is a dynamic clustering sub-region pixel band corresponding to the largest number of abnormal pixel change coordinate points and the largest abnormal deviation value.

[0104] Input the pixel band of the abnormal dynamic clustering sub-region into the Gaussian kernel convolution, diffuse the corresponding abnormal pixel points in the pixel band of the abnormal dynamic clustering sub-region, and obtain the continuous abnormal pixel point coordinate space;

[0105] Feeding back the continuous abnormal pixel point coordinate space to the feature extraction layer to perform secondary correction on the liquid level line position prediction error loss and the predicted liquid level line coordinate space to obtain the corrected liquid level line position prediction error loss and the predicted liquid level line coordinate space;

[0106] The infusion bag segmentation area space to be enhanced is constructed based on the corrected liquid level line position prediction error loss, the predicted liquid level line coordinate space, and the infusion bag segmentation area feature space.

[0107] Based on the available bandwidth, packet loss rate, and transmission delay corresponding to the number of images transmitted simultaneously in different time periods in the network transmission status, the network quality evaluation layer is used to construct the network quality index corresponding to different time periods.

[0108] Based on the network quality index corresponding to different time periods and the preset network quality level interval, the network transmission level corresponding to different time periods is obtained;

[0109] Furthermore, in this embodiment, the preset network quality level interval is specifically as follows: preset NQI thresholds (excellent ≥ 0.8, good ≥ 0.6, fair ≥ 0.4, poor < 0.4), and dynamic matching of image resolution strategies; wherein NQI is the abbreviation of network quality index;

[0110] The network transmission levels corresponding to the different time periods and the preset image enhancement resolution table are input into the hierarchical correction layer to simulate the enhanced transmission of the segmented area space of the infusion bag to be enhanced, so as to obtain the transmission delay and packet loss of each resolution image at different network transmission levels and the corresponding network level-image quality mapping matrix.

[0111] Based on the different pixel spaces of the masked infusion bag segmentation areas and the predicted liquid level position combined with the real-time infusion rate, a probability evaluation layer with a built-in Bayesian function is used to obtain the replacement delay risk probability value and probability evaluation loss corresponding to each infusion bag at the current moment.

[0112] The replacement delay risk probability value corresponding to each infusion bag at the current moment is used as the processing priority of the infusion bag segmentation area space to be enhanced corresponding to each infusion bag captured image;

[0113] Furthermore, in this embodiment, the processing priority is set accordingly according to the replacement delay risk probability value. The larger the replacement delay risk probability value, the higher the corresponding processing priority. Furthermore, the processing priority in this embodiment is the priority for image enhancement and transmission of the infusion bag image, that is, the larger the corresponding infusion bag replacement delay risk probability value, the first time the image corresponding to the infusion bag is enhanced and transmitted to the corresponding nurse PAD for early warning.

[0114] Feeding back the processing priority of the space of the segmented area of the infusion bag to be enhanced corresponding to each infusion bag captured image into the simulated enhancement transmission process, simulating the optimal image quality corresponding to each processing priority image under different network levels, and using transmission delay as a measurement indicator to construct a processing priority-network level-image quality-transmission delay mapping matrix space;

[0115] The processing priority-network level-image quality-transmission delay mapping matrix space is built into the output layer. The infusion bag segmentation area space to be enhanced, the network quality index, and the processing priority corresponding to each infusion bag at the current moment are input into the output layer to obtain the infusion bag image of the corresponding resolution image quality and the corresponding image enhancement loss;

[0116] The loss function of the transmission correction model is constructed based on the corrected liquid level position prediction error loss, probability assessment loss, transmission delay and packet loss loss, and image enhancement loss. The transmission correction model is simulated and trained, and the transmission delay corresponding to each resolution image quality under different network quality and processing priority is analyzed.

[0117] When the transmission delay corresponding to each resolution image quality is less than the preset transmission delay, the trained transmission correction model is obtained, and the enhanced infusion bag image and the corresponding enhanced resolution at the corresponding network level are output.

[0118] Through multi-dimensional technology collaboration and innovation, significant performance improvements have been achieved in the real-time, accuracy, and resource optimization aspects of infusion monitoring. First, the U-Net architecture based on deep separable convolution is combined with a multi-level feature fusion mechanism. During the encoding stage, a stride-size 2 convolution is used to achieve progressive downsampling, effectively compressing redundant information while retaining the morphological characteristics of the infusion bag. During the decoding process, the SE attention mechanism and cross-level feature splicing solve the feature misalignment problem of traditional segmentation models in complex backgrounds, significantly improving the accuracy of infusion bag regional segmentation and laying a solid foundation for liquid level positioning. Secondly, to address the interference of liquid surface fluctuations in dynamic infusion scenarios, multi-scale pyramid gradient analysis and dynamic mask generation technology are used. Through three-level Gaussian kernel decomposition, multi-level fuzzy processing with σ=1 to 3 is achieved, capturing local details and global deformation features at different scales. Combined with gradient direction filtering and channel attention weighted attenuation mechanism, it effectively suppresses the influence of noise such as reflections and bubbles on liquid level detection. At the same time, the mask continuity is optimized through closing operations and connected domain analysis, reducing the liquid level coordinate prediction error by approximately 40%. Furthermore, through optical flow analysis and dynamic clustering sub-region detection technology, a dense optical flow field is constructed and an interquartile range statistical model is used to identify clusters of abnormally moving pixels. This, combined with the Gaussian kernel diffusion algorithm, dynamically corrects the deformation characteristics of infusion bags, improving the system's robustness to sudden changes in infusion rate or external interference by approximately 35%. Regarding network transmission optimization, an innovative dynamic assessment model based on the Network Quality Index (NQI) is introduced, mapping bandwidth, packet loss rate, and latency parameters to network transmission levels. Combined with a processing priority-network level-image quality mapping matrix, this allows for adaptive adjustment of image resolution and enhancement strategies, ensuring that critical infusion bag monitoring data is prioritized while further improving network resource utilization. Furthermore, a Bayesian function-based replacement delay risk probability model integrates multiple data sources, including fluid volume, flow rate, and network status, in real time to construct a dynamic risk assessment framework, significantly improving early warning accuracy. A processing priority feedback mechanism optimizes task scheduling logic, ensuring that enhanced annotation information for high-risk infusion bags is transmitted to the nursing terminal via the optimal path. The synergistic effect of the above technical features enables the system to achieve systematic optimization of fluid volume monitoring accuracy, abnormal response speed and resource allocation efficiency in complex clinical environments, providing a scalable technical paradigm for improving the level of infusion safety management.

[0119] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific implementation methods. The above-mentioned specific implementation methods are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also change, modify, replace and modify the above-mentioned embodiments without departing from the scope of protection of the purpose of the present invention and the claims, and all of these are protected by the present invention.

Claims

1. Medical intelligent infusion management system based on visual recognition, characterized by: include: Monitoring module and transmission discrimination module; the monitoring module includes a distributed monitoring unit and a transmission enhancement unit; The distributed monitoring unit collects synchronous images, user information, and network transmission status of at least one or more infusion bags. Based on the number of synchronously collected images, the network transmission status, and the infusion rate, combined with the transmission correction model built into the transmission enhancement unit, a synchronous transmission-corrected infusion bag image set is obtained and the remaining liquid volume and remaining infusion time are marked. Based on the corrected infusion bag image set and user information, the transmission discrimination module is used to obtain the infusion bag replacement delay risk probability; Feedback the delay risk probability of the infusion bag replacement to the transmission enhancement unit, adjust the transmission image enhancement sequence in real time and mark the delay risk probability, and display the enhanced and marked image on the PAD in real time; The construction process of the transmission correction model includes: First, the background segmentation of the infusion bag image is performed through the background segmentation layer to obtain the infusion bag segmentation area set and the background segmentation area set, and the coordinates of the liquid surface line in the infusion bag segmentation area are marked; secondly, based on the infusion bag shape attribute information, the infusion bag segmentation area, the full liquid volume and the initial liquid level height, the three-dimensional model of the infusion bag is obtained through the three-dimensional modeling layer; based on the three-dimensional model of the infusion bag combined with different infusion rates, the liquid surface separation line position function under different infusion rates is fitted through the simulation algorithm combined with the support vector machine kernel function to obtain the separation line position function corresponding to different infusion rates; thirdly, the separation lines in the infusion bag segmentation area are masked, and each masked infusion bag segmentation area and separation line position function are input into the transmission The feature extraction layer of the correction model obtains the feature space of each infusion bag segmentation area and the liquid level coordinate space predicted by the liquid separation line position function; according to the feature space of the infusion bag segmentation area and the liquid level coordinate space predicted by the liquid separation line position function combined with the marked liquid level coordinates, the liquid level position prediction error loss is obtained; fourthly, the continuous infusion bag image sequence constructed by the masked infusion bag segmentation area corresponding to the continuous time points is input into the optical flow layer, and the motion vector of each pixel corresponding to the continuous time points of the masked infusion bag segmentation area is analyzed to obtain the dense optical flow field corresponding to the continuous masked infusion bag segmentation area; based on the change value of the adjacent pixels in the dense optical flow field corresponding to each masked infusion bag segmentation area, the clustering algorithm is combined with the image sequence to obtain the liquid level position prediction error loss; fourthly, the image sequence of the continuous infusion bag segmentation area corresponding to the masked infusion bag segmentation area is input into the optical flow layer, and the motion vector of each pixel corresponding to the continuous time points of the masked infusion bag segmentation area is analyzed to obtain the dense optical flow field corresponding to the continuous masked infusion bag segmentation area; based on the change value of the adjacent pixels in the dense optical flow field corresponding to the masked infusion bag segmentation area, the clustering algorithm is combined with the image sequence to obtain the liquid level position prediction error loss; fourthly, the image sequence of the continuous infusion bag segmentation area corresponding to the masked infusion bag segmentation area is input into the optical flow layer, and the motion vector of the continuous pixel The third quantile is used to construct the dynamic clustering sub-region pixel band space corresponding to each masked infusion bag segmentation area; the pixel change abnormality threshold is constructed according to the pixel change value in the dynamic clustering sub-region pixel band space and the pixel change mean corresponding to the full-load infusion bag, and the abnormal dynamic clustering sub-region pixel band is obtained; the abnormal dynamic clustering sub-region pixel band is the dynamic clustering sub-region pixel band corresponding to the largest number of pixel change abnormal coordinate points and the largest abnormal deviation value; fifth, the abnormal dynamic clustering sub-region pixel band is input into the Gaussian kernel convolution, and the corresponding abnormal pixel points in the abnormal dynamic clustering sub-region pixel band are diffused to obtain a continuous abnormal pixel point coordinate space; the continuous abnormal pixel point coordinate space is fed back to the The feature extraction layer performs secondary correction on the liquid level line position prediction error loss and the predicted liquid level line coordinate space to obtain the corrected liquid level line position prediction error loss and the predicted liquid level line coordinate space; based on the corrected liquid level line position prediction error loss, the predicted liquid level line coordinate space, and the infusion bag segmentation area feature space, the infusion bag segmentation area space to be enhanced is constructed; at the same time, according to the available bandwidth for transmission in different time periods in the network transmission status, the packet loss rate and transmission delay corresponding to the number of synchronously transmitted images, the network quality index corresponding to different time periods is constructed through the network quality assessment layer; based on the network quality index corresponding to different time periods and the preset network quality level interval, the network transmission level corresponding to different time periods is obtained;The network transmission level corresponding to the different time periods and the preset image enhancement resolution table are input into the hierarchical correction layer to simulate the enhanced transmission of the infusion bag segmentation area space to be enhanced, and the transmission delay and packet loss loss of each resolution image at different network transmission levels and the corresponding network level-image quality mapping matrix are obtained; Sixth, based on the different infusion bag segmentation area pixel spaces and predicted liquid level positions after masking combined with the real-time infusion rate, the replacement delay risk probability value and probability evaluation loss corresponding to each infusion bag at the current moment are obtained through the probability evaluation layer with a built-in Bayesian function; the replacement delay risk probability value corresponding to each infusion bag at the current moment is used as the processing priority of the infusion bag segmentation area space to be enhanced corresponding to each infusion bag acquisition image; the processing priority of the infusion bag segmentation area space to be enhanced corresponding to each infusion bag acquisition image is fed back to the simulated enhanced transmission process, simulating the optimal image quality corresponding to each processing priority image under different network levels, and Using transmission delay as a measurement metric, a processing priority-network level-image quality-transmission delay mapping matrix space is constructed. Seventh, the processing priority-network level-image quality-transmission delay mapping matrix space is embedded in the output layer. The current infusion bag segmentation area to be enhanced, the network quality index, and the processing priority corresponding to each infusion bag are input into the output layer to obtain an infusion bag image of the corresponding resolution image quality and the corresponding image enhancement loss. A transmission correction model loss function is constructed using the corrected liquid level position prediction error loss, probability assessment loss, transmission delay and packet loss loss, and image enhancement loss. The transmission correction model is simulated and trained, and the transmission delay corresponding to each resolution image quality under different network qualities and processing priorities is analyzed. When the transmission delay corresponding to each resolution image quality is less than the preset transmission delay, the trained transmission correction model is obtained, and the enhanced infusion bag image and the corresponding enhanced resolution at the corresponding network level are output.

2. The medical intelligent infusion management system based on visual recognition according to claim 1, characterized in that: The process of masking the liquid separation line in the segmentation area of the infusion bag comprises: Based on the segmented region of the infusion bag, a multi-level pyramid feature map corresponding to the segmented region of the infusion bag is generated by a Gaussian pyramid decomposition algorithm combined with a preset number of layers; Based on the multi-level pyramid feature map, the horizontal gradient and vertical gradient convolution are performed on each level of the pyramid feature map through the Sobel operator convolution to obtain the horizontal and vertical gradient amplitudes and gradient direction angles corresponding to each level of the pyramid feature map; Based on the gradient direction angle and the preset gradient direction angle filtering interval, the horizontal and vertical gradient amplitudes are filtered to obtain the filtered horizontal and vertical gradient amplitudes corresponding to each level of the pyramid feature map; The horizontal and vertical gradient amplitudes corresponding to the filtered feature maps of each level of the pyramid are weighted attenuated through the channel convolution attention algorithm, and the feature maps of each level of the pyramid completed by weighted attenuation are upsampled and fused through bilinear interpolation to obtain a multi-level pyramid fusion gradient map.

3. The medical intelligent infusion management system based on visual recognition according to claim 2, characterized in that: The process of masking the liquid separation line in the infusion bag segmentation area further includes: Based on the multi-level pyramid fusion gradient image, the mean and variance of the fused gradient image are obtained through statistical algorithms; The mask interval is set based on the mean and variance corresponding to the fused gradient map, and the initial binary mask corresponding to the liquid level line in the infusion bag segmentation area is generated according to the set mask interval combined with non-maximum suppression; Based on the initial binary mask corresponding to the liquid level line in the infusion bag segmentation area, the binary mask corresponding to the liquid level line in the infusion bag segmentation area is obtained through the 3×3 rectangular structure element S combined with the connected domain filter. The liquid level line marked in the input infusion bag segmentation area is then binary masked based on the binary mask corresponding to the liquid level line in the liquid bag segmentation area.

Images