ICU catheter dislodgement monitoring system based on image processing

Abstract

The application discloses an ICU catheter shedding monitoring system based on image processing, which comprises an ICU catheter image acquisition module, a catheter state visual extraction module, a shedding feature grouping construction module, a reference mode learning module, a local shedding risk coupling network construction module, a shedding risk topology convergence module, a catheter shedding risk discrimination network training module, a visual feature contribution dynamic calculation module and an ICU catheter shedding monitoring module. The application belongs to the field of image processing, and the scheme introduces time consistency constraint to learn the continuous evolution relationship between catheter states, reduces the influence of short-time disturbance on risk discrimination, constructs a local risk network corresponding to multi-vision feature grouping, dynamically update the risk sensitive weight of each visual feature grouping, combines the risk prediction entropy to dynamically adjust the contribution degree of each feature grouping, suppresses the influence of short-time disturbance and enhances the evolution consistency across patient states, and thus improves the generalization ability of catheter shedding risk discrimination.

Description

Technical Field

[0001] This invention relates to the field of image processing, specifically to an ICU catheter dislodgement monitoring system based on image processing. Background Technology

[0002] An ICU catheter dislodgement monitoring system is an intelligent monitoring system that automatically identifies catheter position changes, loosening, or dislodgement risks by real-time monitoring and analysis of images or videos of the catheter area in ICU patients, and issues early warnings. However, conventional ICU catheter dislodgement monitoring systems suffer from problems such as strong features masking weak features, missing some dislodgement signs, and being easily affected by instantaneous changes caused by nursing operations, leading to a high false alarm rate. Conventional ICU catheter dislodgement monitoring systems also suffer from dominant information bias, which can easily lead to different dislodgement signs interfering with each other in the complex nursing environment of the ICU. Summary of the Invention

[0003] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an image processing-based ICU catheter dislodgement monitoring system. Addressing the problems of strong features masking weak features, missing some dislodgement signs, and susceptibility to instantaneous changes caused by nursing procedures, leading to high false alarm rates in conventional ICU catheter dislodgement monitoring systems, this solution constructs multiple groups of visual features to learn the evolution patterns of different catheter dislodgement signs, avoiding interference between different visual features. Subsequently, a reference pattern library is established, using reference pattern attribution relationships to uniformly represent different catheter states, forming a stable state representation. Based on the reference pattern attribution matrix, a local dislodgement risk coupling network is constructed, and temporal consistency constraints are introduced to learn the continuous evolution relationship between catheter states, reducing the impact of nursing procedures and patient error. The impact of short-term disturbances such as body movement and lighting changes on risk assessment; addressing the problem that conventional ICU catheter dislodgement monitoring systems suffer from dominant information bias, which easily leads to mutual interference between different dislodgement signs in the complex nursing environment of the ICU, resulting in poor accuracy of dislodgement monitoring, this solution constructs a local risk network corresponding to multiple visual feature groups and uses a weighted aggregation method to form a global risk topology; through the dynamic updating of the risk-sensitive weights of each visual feature group, a two-way feedback learning mechanism is formed between the topology structure and feature contribution, and the contribution of each feature group is dynamically adjusted in combination with risk prediction entropy. This enables adaptive highlighting of key dislodgement signs, suppression of short-term interference effects, and enhancement of consistency across patient state evolution in real ICU monitoring scenarios, thereby improving the generalization ability of catheter dislodgement risk assessment.

[0004] The technical solution adopted by the present invention is as follows: The ICU catheter dislodgement monitoring system based on image processing provided by the present invention includes an ICU catheter image acquisition module, a catheter status visual extraction module, a dislodgement feature grouping and construction module, a reference pattern learning module, a local dislodgement risk coupling network construction module, a dislodgement risk topology aggregation module, a catheter dislodgement risk discrimination network training module, a visual feature contribution dynamic calculation module, and an ICU catheter dislodgement monitoring module.

[0005] The ICU catheter image acquisition module acquires historical ICU patient catheter area image sequences and performs grayscale normalization processing;

[0006] The catheter state visual extraction module extracts texture response features, edge stability features, tube geometric features, fixed region deformation features, and temporal motion features based on standardized catheter images.

[0007] The shedding feature grouping construction module forms feature combinations corresponding to different catheter state evolution modes through visual grouping based on the visual features output by the catheter state visual extraction module.

[0008] The reference pattern learning module establishes a reference pattern set for each visual feature group, and obtains the reference pattern attribution matrix corresponding to the duct image sample by optimizing the attribution relationship between visual features and reference patterns.

[0009] The local detachment risk coupling network construction module constructs a local detachment risk coupling network based on the reference pattern attribution matrix; and establishes the local risk coupling relationship between the duct states within each visual feature group.

[0010] The detachment risk topology convergence module performs weighted fusion and alternating optimization on the local detachment risk coupling network corresponding to the visual feature grouping to generate a global risk topology matrix.

[0011] The catheter dislodgement risk discrimination network training module uses the global risk topology matrix as the graph structure adjacency relationship and the reference pattern attribution vector as the node feature to construct and train a graph convolutional catheter dislodgement risk discrimination network.

[0012] During the training of the risk discrimination network, the visual feature contribution dynamic calculation module dynamically evaluates and updates the weights of each visual feature group based on the output of the global risk topology matrix and the risk prediction entropy.

[0013] The ICU catheter dislodgement monitoring module uses a trained catheter dislodgement risk discrimination network to monitor catheter dislodgement from real-time acquired images.

[0014] Furthermore, the ICU catheter image acquisition module uses bedside ICU cameras to continuously acquire historical ICU patient catheter area image sequences, labels the historical catheter images with catheter dislodgement risk tags, and standardizes the grayscale of the catheter images.

[0015] Furthermore, the duct state visual extraction module establishes a multi-source visual system: it uses multi-source visual features to collaboratively represent the duct state; and constructs five types of feature vectors: texture response features, edge stability features, tube geometric features, fixed region deformation features, and temporal motion features.

[0016] Furthermore, the detachment feature grouping construction module constructs detachment feature groups based on the catheter state feature vector output by the catheter state visual extraction module; the visual groups are generated in a random grouping manner, each group contains at least one visual feature, and each visual feature appears in at least one visual feature group.

[0017] Furthermore, the reference pattern learning module constructs a reference pattern library, extracts reference patterns that can represent typical ductal states, and maps samples with similar ductal states to similar reference patterns, thereby providing a stable state representation for the subsequent construction of the detachment risk coupling network and obtaining the reference pattern attribution matrix.

[0018] Furthermore, the local detachment risk coupling network construction module constructs a local detachment risk coupling network based on the reference pattern attribution matrix output by the reference pattern learning module; a local detachment risk coupling network is established within each group of visual features, and a state evolution structure is constructed by learning the coupling relationship between different duct states, thereby extracting the potential evolution law in the duct detachment process; a time consistency constraint is introduced in the learning process of the local detachment risk coupling network, so that the local detachment risk coupling network not only maintains the reference pattern attribution structure, but also maintains the continuous evolution law of the duct state in the time dimension.

[0019] Furthermore, the detachment risk topology aggregation module constructs a unified detachment risk topology; it aggregates all local risk networks to form a global state evolution structure.

[0020] An alternating optimization strategy is used to iteratively optimize the global risk topology matrix and the risk-sensitive weights of visual feature grouping; the projection gradient optimization algorithm is used to iteratively solve the optimization objective to obtain the global risk topology matrix.

[0021] Furthermore, the visual feature contribution dynamic calculation module dynamically updates the risk-sensitive weights of each visual feature group based on the global risk topology matrix; it also dynamically updates the risk-sensitive weights of each visual feature group according to the consistency between each local detachment risk coupling network and the global risk topology, and feeds the updated risk-sensitive weights back to the detachment risk topology aggregation module to participate in the next round of global risk topology matrix optimization, thereby achieving collaborative optimization of the visual feature group risk-sensitive weights and the global risk topology matrix; and it introduces risk prediction entropy to construct risk-sensitive factors, and dynamically adjusts the risk-sensitive weights corresponding to the visual feature groups according to the prediction stability.

[0022] Furthermore, the catheter dislodgement risk discrimination network training module uses the global risk topology matrix as input to establish a graph structure risk discrimination network, and learns the mapping relationship between the catheter state evolution structure and the dislodgement risk;

[0023] A global risk topology matrix is ​​used as the graph adjacency relation, and the reference mode attribution vector of the corresponding catheter image sample is used as the node feature. Both are input into the graph convolutional risk discrimination network. The global risk topology matrix describes the state evolution coupling relationship between catheter image samples, and the reference mode attribution vector describes the catheter state representation of the catheter image sample itself. Both are input into the graph convolutional risk discrimination network, and the output is the category judgment result of the four states corresponding to the ICU catheter image.

[0024] Furthermore, the ICU catheter dislodgement monitoring module uses a trained catheter dislodgement risk discrimination network to perform online detection on real-time acquired catheter images, outputs the catheter dislodgement risk category and generates an early warning result. It adopts a continuous judgment mechanism: when consecutive frames of real-time catheter images are all judged to be in a high-risk dislodgement state or a dislodgement state, a formal alarm is triggered; otherwise, real-time detection continues.

[0025] The beneficial effects achieved by adopting the above solution are as follows:

[0026] (1) To address the problems of strong features masking weak features in conventional ICU catheter dislodgement monitoring systems, which lead to the omission of some dislodgement signs and are easily affected by instantaneous changes caused by nursing operations, resulting in a high false alarm rate, this solution constructs multiple groups of visual features to learn the evolution patterns of different catheter dislodgement signs, thereby avoiding mutual interference between different visual features; then, a reference pattern library is established to uniformly represent different catheter states by reference pattern attribution relationships, forming a stable state expression; a local dislodgement risk coupling network is constructed based on the reference pattern attribution matrix, and time consistency constraints are introduced to learn the continuous evolution relationship between catheter states, thereby reducing the impact of short-term disturbances such as nursing operations, patient movement, and light changes on risk judgment.

[0027] (2) In view of the problem that conventional ICU catheter dislodgement monitoring systems have a dominant information bias, which can easily lead to different dislodgement signs interfering with each other in the complex nursing environment of the ICU, resulting in poor accuracy of dislodgement monitoring, this solution constructs a local risk network corresponding to multiple visual feature groups and uses a weighted aggregation method to form a global risk topology. Through the dynamic update of the risk sensitivity weight of each visual feature group, the topology structure and feature contribution form a two-way feedback learning mechanism. Combined with the risk prediction entropy, the contribution of each feature group is dynamically adjusted, which can adaptively highlight key dislodgement signs, suppress short-term interference effects and enhance the consistency of cross-patient state evolution in the real monitoring scenario of the ICU, thereby improving the generalization ability of catheter dislodgement risk judgment. Attached Figure Description

[0028] Figure 1 A schematic diagram of the ICU catheter dislodgement monitoring system based on image processing provided by the present invention.

[0029] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0030] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0031] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0032] Example 1, see Figure 1 The ICU catheter dislodgement monitoring system based on image processing provided by the present invention includes an ICU catheter image acquisition module, a catheter status visual extraction module, a dislodgement feature grouping and construction module, a reference pattern learning module, a local dislodgement risk coupling network construction module, a dislodgement risk topology aggregation module, a catheter dislodgement risk discrimination network training module, a visual feature contribution dynamic calculation module, and an ICU catheter dislodgement monitoring module.

[0033] The ICU catheter image acquisition module acquires historical ICU patient catheter area image sequences and performs grayscale normalization processing;

[0034] The catheter state visual extraction module extracts texture response features, edge stability features, tube geometric features, fixed region deformation features, and temporal motion features based on standardized catheter images.

[0035] The shedding feature grouping construction module forms feature combinations corresponding to different catheter state evolution modes through visual grouping based on the visual features output by the catheter state visual extraction module.

[0036] The reference pattern learning module establishes a reference pattern set for each visual feature group, and obtains the reference pattern attribution matrix corresponding to the duct image sample by optimizing the attribution relationship between visual features and reference patterns.

[0037] The local detachment risk coupling network construction module constructs a local detachment risk coupling network based on the reference pattern attribution matrix; and establishes the local risk coupling relationship between the duct states within each visual feature group.

[0038] The detachment risk topology convergence module performs weighted fusion and alternating optimization on the local detachment risk coupling network corresponding to the visual feature grouping to generate a global risk topology matrix.

[0039] The catheter dislodgement risk discrimination network training module uses the global risk topology matrix as the graph structure adjacency relationship and the reference pattern attribution vector as the node feature to construct and train a graph convolutional catheter dislodgement risk discrimination network.

[0040] During the training of the risk discrimination network, the visual feature contribution dynamic calculation module dynamically evaluates and updates the weights of each visual feature group based on the output of the global risk topology matrix and the risk prediction entropy.

[0041] The ICU catheter dislodgement monitoring module uses a trained catheter dislodgement risk discrimination network to monitor catheter dislodgement from real-time acquired images.

[0042] Example 2, see Figure 1 This embodiment is based on the above embodiment. The ICU catheter image acquisition module uses a bedside ICU camera to continuously acquire historical ICU patient catheter area image sequences, retaining catheter status change information, and labeling historical catheter images with catheter dislodgement risk tags, including normal fixation, slight displacement, high-risk dislodgement, and dislodged states. To eliminate the influence of differences in camera exposure intensity, the grayscale of the catheter images is standardized. The standardization is expressed as follows: ; It is the grayscale value of the duct image after time t is normalized; These are the coordinates of the catheter image; The grayscale value of the original catheter image at time t; It is the average grayscale value of the current catheter image; It is the current standard deviation of the grayscale of the catheter image; when Less than the preset threshold (Minimum positive number), let ;

[0043] After standardization, the grayscale distribution of the image is transformed into a zero-mean, unit-variance form, eliminating brightness shifts among different patients and under different lighting conditions, and improving the stability of subsequent feature extraction.

[0044] Example 3, see Figure 1 Based on the above embodiments, the catheter status visual extraction module establishes a multi-source visual system to avoid the inability of a single feature to fully characterize the catheter status: catheter detachment is usually accompanied by various visual signs such as changes in catheter contour, deformation of the fixation film, displacement of the catheter outlet, and enhanced local motion. Different signs correspond to different physical change processes. Therefore, multi-source visual features are used to collaboratively characterize the catheter status.

[0045] Construct five types of feature vectors:

[0046] Texture response features are derived using a convolutional coding network. The network structure is: Conv3×3-32 → Conv3×3-64 → Conv3×3-128 → GlobalAveragePooling → 128-dimensional feature vectors; represented as: ; It is the texture response feature vector of the i-th frame image; It is a convolutional coding network; It is a standardized image of the catheter;

[0047] Edge stability characteristics were analyzed, and the Sobel gradient operator was used to calculate the edge response intensity of the duct profile. ;in, It is the horizontal grayscale gradient; It is the gray-level gradient in the vertical direction;

[0048] The geometric features of the catheter were analyzed, and the standardized catheter image was segmented using a catheter segmentation network (U-Net) to obtain a catheter region mask image. The number of catheter pixels was then counted based on the catheter region mask. ; This is the duct region mask in the i-th frame; H and W are the image height and width, respectively; after obtaining the duct region mask, the MedialAxisTransform algorithm is used for skeleton extraction to obtain the duct center skeleton, and the duct centerline length is calculated based on the coordinates of the skeleton points. ; and These are the coordinates of the p-th and p+1-th skeleton points, respectively; This represents the total number of skeletal points on the catheter central skeleton; and the local curvature of each skeletal point is calculated based on the catheter central skeleton curve. A larger curvature value indicates a more pronounced degree of local bending of the catheter; and It is the first derivative of the skeleton curve at the i-th sampling point; and It is the second derivative of the skeleton curve at the i-th sampling point; further, the overall average curvature of the catheter is calculated. n is the number of skeleton sampling points involved in curvature calculation;

[0049] Deformation characteristics of a fixed region , This indicates the rate of change of the area covered by the film; This is the area of ​​the current screen protector application area; This is the initial film application area; the film application area is segmented using U-Net to obtain the film application area mask. Then, the area of ​​the film application area was calculated. The initial film application area was calculated from the baseline image after the catheter was fixed.

[0050] Temporal motion characteristics, , It is the average motion intensity in the ductal region; and These are the standardized images of the duct region at time i and time ik, respectively, and are two-dimensional grayscale matrices; K is the total number of historical reference times.

[0051] Example 4, see Figure 1 This embodiment is based on the above embodiment. The detachment feature grouping construction module constructs detachment features based on the catheter state feature vector output by the catheter state visual extraction module. The texture response features, edge stability features, tube body geometric features, fixed area deformation features, and temporal motion features obtained by the catheter state visual extraction module reflect different physical properties of the catheter state. The catheter detachment process of different patients often shows different signs of evolution. In order to avoid some key detachment signs being covered by other features, multiple sets of visual feature groups are constructed so that different groups learn different detachment evolution patterns, realizing the transformation from single path recognition to multi-evolution path recognition, and improving the ability to capture detachment signs in complex scenarios.

[0052] set up Represents the complete set of visual features; construct Q groups of feature groups, which are generated randomly, each group containing at least one visual feature, and each visual feature appearing in at least one visual feature group; the q-th visual feature group is represented as T represents texture response features; E represents edge stability features; G represents tube geometry features, consisting of tube area, centerline length, and average curvature; D represents fixed region deformation features; and M represents temporal motion features. It is a visual feature in the qth visual feature group. It is the number of features contained in a visual feature group.

[0053] Example 5, see Figure 1 This embodiment is based on the above embodiment. The reference pattern learning module constructs a reference pattern library and uses a small number of reference patterns to describe a large number of image states: the duct state has continuous change characteristics and there are repeated structures among a large number of image samples. Therefore, reference patterns that can represent typical duct states are extracted, and the sample states are compressed using reference patterns so that samples with similar duct states are mapped to similar reference patterns, thereby providing a stable state representation for the subsequent construction of the detachment risk coupling network.

[0054] set up This represents the set of reference patterns; m is the number of reference patterns.

[0055] For any catheter image sample feature vector, the coupling relationship between the sample and the reference pattern is established as follows: ;constraint ;in, is the feature vector of the i-th catheter image sample; H is the reference mode attribution matrix; It is a penalty coefficient used to control the smoothness of attribution assignment, and its value is

[10] . -4 ,10]; is the degree of attribution of the i-th catheter image to the j-th reference pattern, initialized to 1 / m, and solved using an alternating iterative optimization algorithm. Under the condition of a fixed reference pattern set P, the attribution matrix H is updated. Under the condition of a fixed attribution matrix H, the weighted mean is used to update the new reference pattern set P. This process is repeated until the objective function converges to obtain the final reference pattern attribution matrix; N is the total number of catheter image samples. It is the feature vector (center vector) corresponding to the j-th reference pattern. Collect all duct state feature vectors during the training phase, initialize them with K-Means++ to obtain the initial reference pattern, and then continuously update it during the optimization process to finally obtain the reference pattern library.

[0056] Example 6, see Figure 1This embodiment is based on the above embodiment. The local detachment risk coupling network construction module constructs the local detachment risk coupling network based on the reference pattern attribution matrix output by the reference pattern learning module. Catheter detachment is a dynamic process that evolves continuously from a normal fixed state to a detached state. There are potential evolutionary coupling relationships between different duct states. It is difficult to fully utilize the continuous evolutionary information between states by using only a single frame image for classification. Therefore, a local detachment risk coupling network is established within each group of visual features. By learning the coupling relationship between different duct states, a state evolution structure is constructed, thereby extracting the potential evolutionary laws in the duct detachment process.

[0057] For the q-th visual feature group, let the reference pattern attribution matrix corresponding to the q-th visual feature group be . ; It is the degree to which the i-th duct image sample belongs to the j-th reference pattern in the q-th visual feature group;

[0058] Furthermore, a local detachment risk coupling network is constructed, represented as: ; This represents the decoupling strength between state i and state j; the larger the value, the more similar the decoupling evolution trends of the two states. It is the local detachment risk coupling network corresponding to the qth group of visual features;

[0059] The local detachment risk coupling network is initialized as a random non-negative matrix and the optimization objective is obtained by solving the alternating least squares optimization algorithm. Since ICU catheter detachment is a continuous evolution process and the catheter state at adjacent time points has strong temporal continuity, a time consistency constraint is introduced in the learning process of the local detachment risk coupling network so that the local detachment risk coupling network not only maintains the reference pattern belonging structure, but also maintains the continuous evolution law of the catheter state in the time dimension.

[0060] The change in reference model attribution between adjacent time points is expressed as follows: ; and then construct the reference mode attribution change matrix. ; It is the reference mode attribution vector of the i-th catheter image sample in the q-th visual feature group at the corresponding time. The reference mode attribution vector is the row vector of the reference mode attribution matrix corresponding to a single catheter image sample, which is used to represent the probability distribution of the catheter image sample belonging to each reference mode. It is the reference pattern attribution matrix of the qth visual feature group at the corresponding time.

[0061] The final optimization objective is expressed as: ;

[0062] It is the local risk network smoothing coefficient, with a value of

[10] .-4 ,10]; It is the time consistency constraint coefficient, used to control the degree of influence of the time consistency constraint term on the overall optimization objective, and its value is

[10] . -4 ,10]; It is the structural reconstruction error term, used to constrain the local detachment risk coupling network to maintain the original state distribution structure. If two duct states have similar reference mode attribution relationships, the coupling strength in the local detachment risk coupling network will increase.

[0063] It is a network complexity constraint term used to suppress the generation of abnormal coupling relationships and improve the stability of coupled networks with local detachment risk.

[0064] It is a time consistency constraint term used to constrain the local detachment risk coupling network to maintain the consistency of the evolution trend of the catheter state within adjacent time windows. When the catheter is in a normal fixed state and gradually evolves into a detached state, the change of the reference pattern in adjacent time windows is continuous. Therefore, the time consistency constraint can suppress instantaneous abnormal disturbances caused by nursing operations and light fluctuations.

[0065] By performing the above operations, this solution addresses the problems of conventional ICU catheter dislodgement monitoring systems, such as strong features masking weak features, missing some dislodgement signs, and being easily affected by instantaneous changes caused by nursing operations, leading to a high false alarm rate. This solution constructs multiple groups of visual features to learn the evolution patterns of different catheter dislodgement signs, avoiding interference between different visual features. Subsequently, a reference pattern library is established to uniformly represent different catheter states using reference pattern attribution relationships, forming a stable state representation. Based on the reference pattern attribution matrix, a local dislodgement risk coupling network is constructed, and time consistency constraints are introduced to learn the continuous evolution relationship between catheter states, reducing the impact of short-term disturbances such as nursing operations, patient movement, and changes in lighting on risk assessment.

[0066] Example 7, see Figure 1 This embodiment is based on the above embodiment. The shedding risk topology aggregation module constructs a unified shedding risk topology. Different visual features are grouped to form different local risk networks. Each local risk network only reflects the evolution law of a certain type of shedding symptoms. Therefore, all local risk networks are aggregated to form a global state evolution structure.

[0067] An alternating optimization strategy is used to iteratively optimize the global risk topology matrix and the risk sensitivity weights of visual feature groups; firstly, the risk sensitivity weights of each visual feature group are initialized as follows: During the initialization phase, each visual feature group has the same initial risk sensitivity weight. In subsequent alternating optimization processes, the risk sensitivity weight of each visual feature group is automatically adjusted according to the evolution of the duct state.

[0068] Then, fix the risk-sensitive weights of the visual feature groupings and optimize and update the global risk topology matrix; after obtaining the new global risk topology matrix, fix the global risk topology matrix again and update the risk-sensitive weights of the visual feature groupings; repeat the above two steps until the global risk topology matrix converges and output the final global risk topology matrix.

[0069] The global risk topology matrix is ​​represented as follows: The global risk topology matrix is ​​obtained by iteratively solving the optimization objective using the projection gradient optimization algorithm. ;constraint ; It is the global risk coupling strength between state i and state j; It is the global topological regularization coefficient; and These are the detachment coupling strength vectors corresponding to the i-th and j-th catheter image samples in the local detachment risk coupling network, respectively. The row vector.

[0070] Example 8, see Figure 1 This embodiment is based on the above embodiment. The visual feature contribution dynamic calculation module dynamically updates the risk sensitivity weights of each visual feature group based on the global risk topology matrix obtained during the alternating optimization process of the detachment risk topology convergence module. There are obvious individual differences in the catheter detachment process among different patients, and the ability of different visual feature groups to represent the evolution law of catheter detachment is not consistent. Therefore, under the condition of fixed global risk topology matrix, the risk sensitivity weights of each visual feature group are dynamically updated according to the consistency between each local detachment risk coupling network and the global risk topology. The updated risk sensitivity weights are fed back to the detachment risk topology convergence module to participate in the next round of global risk topology matrix optimization, so as to achieve the collaborative optimization of visual feature group risk sensitivity weights and global risk topology matrix.

[0071] A risk sensitivity factor is constructed by introducing risk prediction entropy. The risk sensitivity weights corresponding to the visual feature groups are dynamically adjusted based on prediction stability. The risk sensitivity weight corresponding to the q-th visual feature group is expressed as: ;in, It is the risk-sensitive weight corresponding to the qth group of visual features; It is the risk prediction entropy corresponding to the q-th group of visual features. , It is the predicted probability that the i-th duct image sample in the q-th visual feature group belongs to the c-th detachment risk state; It is the risk sensitivity adjustment coefficient, with values ​​ranging from [0.5, 1]. It is a risk-sensitive factor. When the risk prediction entropy corresponding to the visual feature group is small, the risk-sensitive factor increases, and the corresponding visual feature group obtains a greater risk-sensitive weight.

[0072] The updated risk-sensitive weights re-participate in the next round of global risk topology matrix optimization, when the global risk topology matrix obtained in two consecutive iterations satisfies The alternating optimization ends at the specified time, and the final global risk topology matrix is ​​output. and These are the global risk topology matrices obtained from rounds f+1 and f, respectively;

[0073] After alternating optimization, visual feature groups that contribute more automatically receive greater convergence weights during the global risk topology construction process, while visual feature groups that contribute less automatically receive smaller weights. This allows the final global risk topology matrix to retain more state evolution information with strong catheter dislodgement characterization capabilities, thereby improving cross-patient generalization ability and the accuracy of catheter dislodgement risk discrimination.

[0074] Example 9, see Figure 1 This embodiment is based on the above embodiment. The catheter dislodgement risk discrimination network training module uses the global risk topology matrix as input to establish a graph structure risk discrimination network and learns the mapping relationship between the catheter state evolution structure and the dislodgement risk.

[0075] A global risk topology matrix is ​​used as the graph adjacency relation, and the reference pattern attribution vector of the corresponding catheter image sample is used as the node feature. Both are input into the graph convolutional risk discrimination network. The global risk topology matrix describes the state evolution coupling relationship between catheter image samples, and the reference pattern attribution vector describes the catheter state representation of the catheter image sample itself. Both are input into the graph convolutional risk discrimination network, enabling the network to simultaneously utilize the catheter state representation and the catheter state evolution relationship to achieve catheter detachment risk discrimination. After the network is trained, the graph convolutional risk discrimination network parameters are saved for direct use in the online monitoring stage. The output is the category judgment result of four states corresponding to ICU catheter images.

[0076] Construct a graph convolutional risk discrimination network;

[0077] The network structure is as follows:

[0078] The first graph convolutional layer: GraphConv(m,128), using ReLU as the activation function;

[0079] The second graph convolutional layer: GraphConv(128,64), using ReLU as the activation function;

[0080] The third graph convolutional layer: GraphConv(64,32), with ReLU activation function;

[0081] Fully connected layer: FC(32,16);

[0082] Output layer: FC(16,4);

[0083] The classification function used is Softmax; the output dimension 4 corresponds to the normal fixed state, the slightly displaced state, the high-risk detachment state, and the detached state.

[0084] The hyperparameters are initialized using Xavier; the loss function is the multi-class cross-entropy loss function, expressed as follows: L is the network loss value; N is the number of training samples; c is the class number; It is the true label of the duct image sample i in category c, when it belongs to category i. ,otherwise ; This represents the probability that the network predicts sample i belongs to category c.

[0085] By performing the above operations, this solution addresses the problem of dominant information bias in conventional ICU catheter dislodgement monitoring systems, which can easily lead to interference between different dislodgement signs in the complex nursing environment of the ICU, resulting in poor accuracy of dislodgement monitoring. Instead, it constructs a local risk network corresponding to multiple visual feature groups and uses a weighted aggregation method to form a global risk topology. Through dynamic updates of the risk-sensitive weights of each visual feature group, a two-way feedback learning mechanism is established between the topology and feature contribution. Combined with the dynamic adjustment of the contribution of each feature group by risk prediction entropy, this solution can adaptively highlight key dislodgement signs, suppress short-term interference, and enhance consistency across patient states in real ICU monitoring scenarios, thereby improving the generalization ability of catheter dislodgement risk assessment.

[0086] Example 10, see Figure 1 This embodiment is based on the above embodiment. The ICU catheter dislodgement monitoring module uses the trained catheter dislodgement risk discrimination network to perform online detection on real-time acquired catheter images, output the catheter dislodgement risk category and generate an early warning result. In actual operation, it continuously acquires real-time catheter images collected by the ICU bedside camera. Following the same data processing flow as the training phase, it sequentially completes image standardization, multi-source visual feature extraction and visual feature grouping construction. Then, it uses the reference pattern library obtained in the training phase to calculate the reference pattern belonging vector corresponding to the current catheter image, and combines it with the trained graph convolutional risk discrimination network to identify the catheter dislodgement risk of the current catheter image. In order to reduce false alarms caused by misjudgment of a single frame image, a continuous judgment mechanism is adopted. When L consecutive frames (values ​​range from 3 to 10) of real-time catheter images are all judged to be in a high-risk dislodgement state or a dislodged state, a formal alarm is triggered; otherwise, real-time detection continues.

[0087] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0088] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An ICU catheter dislodgement monitoring system based on image processing, characterized in that: The system includes an ICU catheter image acquisition module, a catheter status visual extraction module, a dislodgement feature grouping and construction module, a reference pattern learning module, a local dislodgement risk coupling network construction module, a dislodgement risk topology aggregation module, a catheter dislodgement risk discrimination network training module, a visual feature contribution dynamic calculation module, and an ICU catheter dislodgement monitoring module. The ICU catheter image acquisition module acquires historical ICU patient catheter area image sequences and performs grayscale normalization processing; The catheter state visual extraction module extracts texture response features, edge stability features, tube geometric features, fixed region deformation features, and temporal motion features based on standardized catheter images. The shedding feature grouping construction module forms feature combinations corresponding to different catheter state evolution modes through visual grouping based on the visual features output by the catheter state visual extraction module. The reference pattern learning module establishes a reference pattern set for each visual feature group, and obtains the reference pattern attribution matrix corresponding to the duct image sample by optimizing the attribution relationship between visual features and reference patterns. The local detachment risk coupling network construction module constructs a local detachment risk coupling network based on the reference pattern attribution matrix; and establishes the local risk coupling relationship between the duct states within each visual feature group. The detachment risk topology convergence module performs weighted fusion and alternating optimization on the local detachment risk coupling network corresponding to the visual feature grouping to generate a global risk topology matrix. The catheter dislodgement risk discrimination network training module uses the global risk topology matrix as the graph structure adjacency relationship and the reference pattern attribution vector as the node feature to construct and train a graph convolutional catheter dislodgement risk discrimination network. During the training of the risk discrimination network, the visual feature contribution dynamic calculation module dynamically evaluates and updates the weights of each visual feature group based on the output of the global risk topology matrix and the risk prediction entropy. The ICU catheter dislodgement monitoring module uses a trained catheter dislodgement risk discrimination network to monitor catheter dislodgement from real-time acquired images.

2. The ICU catheter dislodgement monitoring system based on image processing according to claim 1, characterized in that: The detachment feature grouping construction module constructs detachment feature groups based on the catheter state feature vector output by the catheter state visual extraction module; the visual groups are generated in a random grouping manner, each group contains at least one visual feature, and each visual feature appears in at least one visual feature group.

3. The ICU catheter dislodgement monitoring system based on image processing according to claim 2, characterized in that: The reference pattern learning module constructs a reference pattern library, extracts reference patterns that can represent typical duct states, and maps samples with similar duct states to similar reference patterns, thereby providing a stable state representation for the subsequent construction of the shedding risk coupling network and obtaining the reference pattern attribution matrix.

4. The ICU catheter dislodgement monitoring system based on image processing according to claim 3, characterized in that: The local detachment risk coupling network construction module constructs a local detachment risk coupling network based on the reference pattern attribution matrix output by the reference pattern learning module. A local detachment risk coupling network is established within each visual feature group. By learning the coupling relationship between different duct states, a state evolution structure is constructed, thereby extracting the potential evolution law in the duct detachment process. A time consistency constraint is introduced in the local detachment risk coupling network learning process, so that the local detachment risk coupling network not only maintains the reference pattern attribution structure, but also maintains the continuous evolution law of the duct state in the time dimension.

5. The ICU catheter dislodgement monitoring system based on image processing according to claim 4, characterized in that: The detachment risk topology aggregation module constructs a unified detachment risk topology; it aggregates all local risk networks to form a global state evolution structure. An alternating optimization strategy is used to iteratively optimize the global risk topology matrix and the risk-sensitive weights of visual feature groupings. The global risk topology matrix is ​​obtained by iteratively solving the optimization objective using the projection gradient optimization algorithm.

6. The ICU catheter dislodgement monitoring system based on image processing according to claim 5, characterized in that: The visual feature contribution dynamic calculation module dynamically updates the risk sensitivity weights of each visual feature group based on the global risk topology matrix. Based on the consistency between each local detachment risk coupling network and the global risk topology, the risk sensitivity weights of each visual feature group are dynamically updated, and the updated risk sensitivity weights are fed back to the detachment risk topology aggregation module to participate in the next round of global risk topology matrix optimization, thereby achieving the collaborative optimization of visual feature group risk sensitivity weights and global risk topology matrix. Furthermore, risk prediction entropy is introduced to construct a risk-sensitive factor, and the risk-sensitive weights corresponding to the visual feature grouping are dynamically adjusted based on the prediction stability.

7. The ICU catheter dislodgement monitoring system based on image processing according to claim 6, characterized in that: The catheter dislodgement risk discrimination network training module uses the global risk topology matrix as input to establish a graph structure risk discrimination network and learns the mapping relationship between the catheter state evolution structure and dislodgement risk; A global risk topology matrix is ​​used as the graph adjacency relation, and the reference pattern attribution vector of the corresponding duct image sample is used as the node feature. Both are input into the graph convolutional risk discrimination network. The global risk topology matrix describes the state evolution coupling relationship between catheter image samples, and the reference mode attribution vector describes the catheter state representation of the catheter image sample itself. Both are input into the graph convolutional risk discrimination network, and the output is the category judgment result of the four states corresponding to the ICU catheter image.

8. The ICU catheter dislodgement monitoring system based on image processing according to claim 7, characterized in that: The ICU catheter image acquisition module uses bedside cameras to continuously acquire historical ICU patient catheter area image sequences, labels the historical catheter images with catheter dislodgement risk tags, and standardizes the grayscale of the catheter images.

9. The ICU catheter dislodgement monitoring system based on image processing according to claim 8, characterized in that: The duct state visual extraction module establishes a multi-source visual system: it uses multi-source visual features to collaboratively represent the duct state; and constructs five types of feature vectors: texture response features, edge stability features, duct geometric features, fixed region deformation features, and temporal motion features.

10. The ICU catheter dislodgement monitoring system based on image processing according to claim 9, characterized in that: The ICU catheter dislodgement monitoring module uses a trained catheter dislodgement risk discrimination network to perform online detection on real-time acquired catheter images, outputs the catheter dislodgement risk category and generates an early warning result. It adopts a continuous judgment mechanism: when consecutive frames of real-time catheter images are all judged to be in a high-risk dislodgement state or a dislodgement state, a formal alarm is triggered; otherwise, real-time detection continues.

Images