A non-invasive blood separation process monitoring method and system based on infrared imaging

Abstract

The application provides a non-invasive blood separation process monitoring method and system based on infrared imaging, and relates to the technical field of medical detection and biological separation, and comprises the following steps: acquiring an infrared image sequence of a blood separation device through an infrared camera, and capturing the thermal radiation distribution characteristics of blood components; adopting a sliding window Kalman filtering algorithm to perform time domain noise reduction on the infrared image sequence, and adaptively adjusting the sliding window length, dynamically updating the noise covariance, and outputting the noise-reduced infrared image; processing the noise-reduced infrared image based on a multi-model switching convolutional neural network architecture, dynamically selecting and loading a convolutional neural network model matched with the current blood separation stage, so as to extract blood separation interface morphological characteristics; and calculating key separation parameters based on the separation interface morphological characteristics, and triggering an early warning signal when the parameters are abnormal. The application realizes non-invasive dynamic monitoring of clinical blood separation.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical detection and biological separation technology, in particular to a non-invasive blood separation process monitoring method and system based on infrared imaging. BACKGROUND

[0002] The statements in this section merely provide background technology related to the present disclosure and do not necessarily constitute prior art.

[0003] Blood separation is the core technology of clinical treatment (such as plasma replacement, component transfusion) and biopharmaceuticals (such as plasma protein purification), and its separation effect directly affects the safety of treatment and the effectiveness of products. The current mainstream monitoring technology has significant limitations: invasive sampling detection requires repeated puncture sampling of separation samples for biochemical analysis, which not only increases the risk of infection for patients, but also destroys the integrity of blood stratification; although radiation imaging technology (such as low-dose X-ray fluoroscopy) can achieve limited visualization, it has cumulative radiation hazards and cannot be used for long-term continuous monitoring; manual experience judgment relies on the subjective evaluation of the operator to assess the clarity of the stratification interface, and cannot quantify key parameters (such as plasma layer height, hematocrit), with an error rate as high as 15%~20%.

[0004] Infrared thermal imaging technology has shown application potential in the field of medical monitoring due to its non-invasive and non-radiation characteristics. However, it has the following technical problems in the context of blood separation:

[0005] (1) Insufficient feature sensitivity: the thermal radiation difference between plasma, red blood cells and platelets is weak (<0.5℃), and environmental temperature fluctuations and device thermal noise result in a signal-to-noise ratio (SNR) of less than 8dB;

[0006] (2) Dynamic response lag: the sedimentation rate of cells in the centrifugal separation process changes nonlinearly, and traditional static image processing methods (such as fixed threshold segmentation) cannot track the migration of the stratification interface in real time;

[0007] (3) Quantitative accuracy defects: existing algorithms (such as edge detection) cannot distinguish the spectral overlap region between plasma and buffer, with an interface positioning error of more than ±3mm, which cannot meet the clinical precision requirements;

[0008] (4) Process monitoring missing: existing technologies focus on measuring and evaluating the static results (volume, purity) of the separation endpoint, and lack dynamic, real-time, quantitative monitoring capabilities for the separation process, which cannot timely detect abnormalities and provide early warnings during the separation process. SUMMARY

[0009] In order to solve at least one technical problem in the background art, in particular, to solve the technical problem that the conventional method cannot non-invasively, in real time and accurately monitor the dynamic process of blood separation, the present application realizes accurate quantitative evaluation of the separation effect through infrared thermal radiation feature capture and dynamic data processing technology. The technical problems of high operation risk, poor real-time performance and weak signal anti-interference ability of the traditional invasive sampling detection and radiation imaging method are solved, providing safe and efficient technical support for clinical blood separation.

[0010] The first aspect of the present application provides an infrared imaging-based non-invasive blood separation process monitoring method, comprising:

[0011] capturing the thermal radiation distribution characteristics of blood components by acquiring an infrared image sequence of the blood separation device through an infrared camera;

[0012] using a sliding window Kalman filter algorithm to perform time domain noise reduction on the infrared image sequence, dynamically updating the noise covariance and adaptively adjusting the sliding window length, and outputting the noise-reduced infrared image;

[0013] processing the noise-reduced infrared image based on a multi-model switching convolutional neural network architecture, dynamically selecting and loading a convolutional neural network model matched with the current blood separation stage to extract blood separation interface morphology features; the convolutional neural network model matched with the current blood separation stage includes a first model optimized for the rapid sedimentation period of red blood cells, a second model optimized for the white blood cell-platelet formation period, and a third model optimized for the blood clarification period;

[0014] calculating key separation parameters based on the separation interface morphology features, and triggering an early warning signal when the parameters are abnormal.

[0015] Further, the dynamic updating of the noise covariance comprises:

[0016] initializing a sliding window with a length of L, and calculating the observation noise variance in real time:

[0017] ;

[0018] wherein, is the i-th frame of original infrared image data, is the Kalman filter prediction value, is the current frame number, is the sliding window length.

[0019] Further, the time domain noise reduction process comprises dynamically switching the filter mode:

[0020] when the signal-to-noise ratio is greater than or equal to the first threshold, a high gain mode is enabled, and the Kalman gain coefficient is increased to accelerate the tracking of separation interface mutations;

[0021] Enable low gain mode when SNR < first threshold, reduce Kalman gain coefficient to suppress ambient noise.

[0022] Further, the key separation parameters include plasma layer height and hematocrit ;

[0023] The plasma layer height is calculated by:

[0024] ;

[0025] The hematocrit is calculated by:

[0026] ;

[0027] wherein, is the plasma area pixel area, is the total pixel area of the container cross section, is the physical height of the container.

[0028] Further, the parameter abnormality trigger condition is:

[0029] or separation interface fluctuation standard deviation ;

[0030] wherein, is the hematocrit warning threshold, is the interface stability threshold.

[0031] Further, the first model is used for the rapid settling phase of red blood cells, and the architecture includes: three layers of convolutional layers, the number of filters is 32, 64 and 64 respectively, and the size is 5x5 pixels, 3x3 pixels and 3x3 pixels respectively; a max-pooling layer; and two layers of fully connected layers, the number of neurons is 128 and 64 respectively;

[0032] The second model is used for the white blood cell-platelet formation phase, and the architecture includes: four layers of convolutional layers, the number of filters is 32, 64, 128 and 128 respectively, and the size is 5x5 pixels, 5x5 pixels, 3x3 pixels and 3x3 pixels respectively; a max-pooling layer; and two layers of fully connected layers, the number of neurons is 256 and 128 respectively;

[0033] The third model is used for the blood clarification phase, and the architecture includes: four layers of convolutional layers, the number of filters is 64, 128, 256 and 256 respectively, and the size is 3x3 pixels; a max-pooling layer; and two layers of fully connected layers, the number of neurons is 512 and 256 respectively;

[0034] The convolutional neural network model matched with the current blood separation stage adopts a ReLU activation function and outputs a separation interface coordinate and a plasma area pixel area. .

[0035] The second aspect of the present application provides a non-invasive blood separation process monitoring system based on infrared imaging, comprising:

[0036] An infrared image acquisition module is configured to acquire an infrared image sequence of a blood separation device through an infrared camera and capture thermal radiation distribution characteristics of blood components.

[0037] A dynamic noise suppression module is configured to perform time domain noise reduction on the infrared image sequence using a sliding window Kalman filtering algorithm, dynamically update noise covariance, and adaptively adjust the sliding window length to output a denoised infrared image.

[0038] A feature enhancement and extraction module is configured to process the denoised infrared image based on a multi-model switching convolutional neural network architecture, dynamically select and load a convolutional neural network model matched with the current blood separation stage to extract blood separation interface morphology features.

[0039] A parameter generation and warning module is configured to calculate key separation parameters based on the separation interface morphology features and trigger a warning signal when the parameters are abnormal.

[0040] The third aspect of the present application provides an electronic device comprising a memory, a processor, and a program stored on the memory and running on the processor, wherein the processor executes the program to implement the steps of the non-invasive blood separation process monitoring method based on infrared imaging according to the first aspect of the present application.

[0041] The fourth aspect of the present application provides a computer readable storage medium having a program stored thereon, wherein the program is executed by a processor to implement the steps of the non-invasive blood separation process monitoring method based on infrared imaging according to the first aspect of the present application.

[0042] The fifth aspect of the present application provides a computer program product comprising software code, wherein the program in the software code executes the steps of the non-invasive blood separation process monitoring method based on infrared imaging according to the first aspect of the present application.

[0043] Compared with the prior art, the non-invasive blood separation process monitoring method and system based on infrared imaging provided by the present application has the following beneficial effects:

[0044] (1) The present application collects an infrared image sequence through an infrared camera, adopts infrared thermal radiation imaging (non-contact, no radiation) to replace invasive puncture or radiation imaging, avoids the risk of puncture infection and the cumulative harm of radiation, and realizes non-invasive and safe monitoring.

[0045] (2) The SW-KF provided by the present application calculates noise variance through a sliding window, can improve the signal-to-noise ratio by 12dB in a weak signal environment, introduces a dynamic multi-model CNN architecture specially designed for the physical characteristics of each stage of blood separation, and realizes the optimal balance between monitoring accuracy and efficiency: in the rapid sedimentation period, the first model realizes the robust positioning of the significant interface with high efficiency; in the buffer layer formation period, the second model captures the fine interface with stronger feature extraction capability; in the clarification period, the third model realizes the accurate judgment of the final state with the strongest abstract ability. This targeted design overcomes the performance limitations of a single model in dealing with changes throughout the process, and can accurately distinguish the spectral overlap region of plasma and buffer, reduce the interface positioning error to <3%, and realize the accurate quantification of the plasma layer height and hematocrit by combining key parameter calculation;

[0046] (3) The present application automatically triggers an early warning (such as an audible and visual prompt) by presetting a hematocrit threshold and an interface fluctuation threshold, replaces manual subjective judgment, reduces operation error, and ensures the stability of the separation process. BRIEF DESCRIPTION OF DRAWINGS

[0047] The drawings accompanying the specification of the present disclosure serve to provide a further understanding of the present disclosure, and the illustrative embodiments of the present disclosure and their descriptions are used to explain the present disclosure, and do not constitute an improper limitation on the present disclosure.

[0048] Figure 1 is a flowchart of the non-invasive blood separation process monitoring method based on infrared imaging provided by the first embodiment of the present application;

[0049] Figure 2 is an example of an infrared image of blood separation in the rapid sedimentation period provided by the first embodiment of the present application;

[0050] Figure 3 is an example of an infrared image of blood separation in the buffer layer formation period provided by the first embodiment of the present application;

[0051] Figure 4 is an example of an infrared image of blood separation in the plasma clarification period provided by the first embodiment of the present application;

[0052] Figure 5 is an abnormal separation early warning interface schematic diagram provided by the first embodiment of the present application;

[0053] Figure 6 is a schematic diagram of a non-invasive blood separation process monitoring system based on infrared imaging provided by the second embodiment of the present application. DETAILED DESCRIPTION

[0054] It should be noted that the following detailed description is exemplary in nature and is intended to provide further description of the application. Unless otherwise defined, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs.

[0055] It is also to be understood that the terminology used herein is for the purpose of describing particular embodiments only and is not intended to be limiting, as the scope of the application will be limited only by the appended claims. Unless defined otherwise, all technical and scientific terms used herein have the same meaning as commonly understood by one of ordinary skill in the art to which this application belongs. Unless otherwise required by context, singular terms shall include pluralities and vice versa. Plural elements can be separated by a hyphenated form of the element name, for example, "multiple components", unless otherwise indicated. Unless otherwise indicated, the use of "or" in the examples herein shall not be understood as an exclusive "or" unless explicitly stated otherwise. Unless otherwise required by context, the use of "in" or "into" shall not be construed as indicating a position within a plane or a surface, but rather shall be understood as indicating a position within a three-dimensional space.

[0056] The embodiments in the application and the features in the embodiments can be combined with each other without conflict.

[0057] All data acquisition of the embodiments is based on compliance with laws and regulations and user consent, and legal application of data.

[0058] Embodiment one

[0059] As Figure 1 , the embodiment provides a non-invasive blood separation process monitoring method based on infrared imaging, comprising:

[0060] An infrared camera is used to collect an infrared image sequence of a blood separation device, and the thermal radiation distribution characteristics of blood components are captured;

[0061] A sliding window Kalman filtering algorithm is used to perform time domain noise reduction on the infrared image sequence, dynamically update the noise covariance, and adaptively adjust the sliding window length, and output the noise-reduced infrared image;

[0062] A multi-model switching-based convolutional neural network architecture is used to process the noise-reduced infrared image, dynamically select and load a convolutional neural network model matched with the current blood separation stage to extract blood separation interface morphology features; the convolutional neural network model matched with the current blood separation stage includes a first model optimized for the rapid sedimentation period of red blood cells, a second model optimized for the white blood cell-platelet formation period, and a third model optimized for the blood clarification period;

[0063] Based on the separation interface morphology features, key separation parameters are calculated, and a warning signal is triggered when the parameters are abnormal.

[0064] The convolutional neural network model provided by the application adopts a multi-model switching architecture, and contains three convolutional neural network models matched with the current blood separation stage, which are optimized for the physical characteristics and visual features of different stages of blood separation. The design is based on the following understanding: blood separation is a dynamic and multi-stage process, and the interface morphology, feature scale and identification difficulty of each stage are significantly different, and it is difficult to achieve optimal accuracy in all stages using a single model. The multi-model switching mechanism dynamically selects and loads the convolutional neural network model that best matches the current blood separation stage according to the time information of the centrifugation process or the interface features of the previous frame.

[0065] Specifically, the dynamic updating of the noise covariance includes:

[0066] Initialize a sliding window with a length of L , and calculate the observation noise variance in real time. The noise caused by the vibration of the centrifuge and the change of the cell sedimentation speed during the blood separation process is time-varying and non-stationary. Fixed parameter filtering cannot adapt. By using the sliding window to estimate the noise in real time, the Kalman filter has adaptive ability, and provides high-quality infrared image data for the next morphological feature extraction.

[0067] The calculation formula of the observation noise variance is as follows:

[0068] ;

[0069] Wherein, is the observation noise variance, is the original infrared image data of the th frame, is the Kalman filter prediction value, is the current frame number, is the sliding window length.

[0070] By initializing the sliding window and calculating the observation noise variance in real time, the noise suppression effect of the Kalman filter is optimized, and the dynamically changing noise environment in the blood separation process is adapted. The problem of low signal-to-noise ratio caused by environmental temperature fluctuations and device thermal noise in infrared images is solved, and the recognition degree of weak signals (such as the small thermal radiation difference between plasma and blood cells) is improved, laying a high-quality data foundation for subsequent feature extraction.

[0071] Specifically, the time domain noise reduction process includes dynamically switching the filtering mode:

[0072] When the signal-to-noise ratio is greater than or equal to the first threshold value (such as 20 dB), it indicates that the system is in the rapid cell sedimentation stage or the signal quality is good, at this time, the high gain mode is enabled, and the Kalman gain coefficient is increased to speed up the tracking of the interface mutation;

[0073] When the signal-to-noise ratio is less than the first threshold value, it indicates that the system is in the late sedimentation stage or in a strong noise interference environment, at which time the low gain mode is enabled to reduce the Kalman gain coefficient to suppress environmental noise. This dual-mode adaptive mechanism ensures the real-time and stability of monitoring during the entire nonlinear separation process. It adapts to the nonlinear changes in the sedimentation rate of cells in centrifugal separation (such as rapid sedimentation at the beginning and slow and stable at the end), avoids tracking lag or excessive smoothing caused by a single filtering mode, and ensures the real-time and stability of interface monitoring.

[0074] In particular, the key separation parameters include the plasma layer height and hematocrit ;

[0075] The calculation formula of the plasma layer height is:

[0076] ;

[0077] The calculation formula of the hematocrit is:

[0078] ;

[0079] wherein, is the pixel area of the plasma region, is the total pixel area of the container cross section, is the physical height of the container.

[0080] The separation effect is quantitatively evaluated, replacing the subjective judgment of artificial, so that the key indicators such as the plasma layer height and the hematocrit can be accurately calculated (error < 3%), providing data-based basis for clinical treatment and biopharmaceuticals.

[0081] In particular, the parameter abnormality triggering condition is:

[0082] or the separation interface fluctuation standard deviation ;

[0083] wherein, is the hematocrit warning threshold, is the interface stability threshold.

[0084] An automated risk prevention and control mechanism is constructed to timely detect separation abnormalities (such as insufficient separation and interface disorder), avoid the omissions of artificial monitoring, and ensure the safety of the treatment or production process.

[0085] Specifically, the first model is used for the rapid settling period of red blood cells, and the architecture thereof comprises: three convolutional layers, the filter numbers of which are 32, 64 and 64 respectively, and the sizes of which are 5*5 pixels, 3*3 pixels and 3*3 pixels respectively; a max-pooling layer; and two fully connected layers, the neuron numbers of which are 128 and 64 respectively.

[0086] The second model is used for the white blood cell-platelet formation period, and the architecture thereof comprises: four convolutional layers, the filter numbers of which are 32, 64, 128 and 128 respectively, and the sizes of which are 5*5 pixels, 5*5 pixels, 3*3 pixels and 3*3 pixels respectively; a max-pooling layer; and two fully connected layers, the neuron numbers of which are 256 and 128 respectively.

[0087] The third model is used for the blood clarification period, and the architecture thereof comprises: four convolutional layers, the filter numbers of which are 64, 128, 256 and 256 respectively, and the sizes of which are all 3*3 pixels; a max-pooling layer; and two fully connected layers, the neuron numbers of which are 512 and 256 respectively.

[0088] The convolutional neural network models matched with the current blood separation stage all adopt ReLU activation functions and output separation interface coordinates and plasma area pixel areas. .

[0089] Working principle of the first model (for the rapid settling period of red blood cells)

[0090] 1. Model architecture:

[0091] The first model is specially designed for the rapid settling period of red blood cells. The architecture thereof comprises, in sequence: an input layer (a 128*128-pixel grayscale image), a convolutional layer 1 (using 32 5*5-pixel filters), a convolutional layer 2 (using 64 3*3-pixel filters), a convolutional layer 3 (using 64 3*3-pixel filters), a max-pooling layer (a 2*2-pixel window, with a step of 2), a fully connected layer 1 (128 neurons), a fully connected layer 2 (64 neurons), and an output layer (3 neurons for coordinates and area).

[0092] 2. Model features and beneficial effects:

[0093] The first model adopts a relatively shallow network depth (3 convolutional layers) and a relatively large initial convolutional kernel (5*5 pixels). Such a design enables the model to have a large receptive field at the shallow layer, so as to quickly capture the large-range and continuous interface features between the red blood cell layer and the upper components, without the need for deep-level complex feature combination.

[0094] The relatively small number of filters (32 -> 64 -> 64) and smaller fully connected layers (128 neurons -> 64 neurons) result in fewer model parameters and high computational efficiency, meeting the high-frequency demand for real-time monitoring during the rapid settling period.

[0095] The first model focuses on solving the technical problem of target prominence during blood sedimentation, avoiding the computational redundancy and overfitting risk that may be caused by using complex models.

[0096] 3. Coordination with sedimentation characteristics:

[0097] For example Figure 2 As shown in the infrared image of blood separation during the rapid settling period of red blood cells, the core characteristics of the rapid settling period of red blood cells are that the proportion of sedimenting substances (red blood cells) is large, the settling speed is fast, and the interface formed is macroscopic and prominent.

[0098] The architecture of the first model is designed to match the characteristics of this period:

[0099] First, large convolution kernels correspond to large targets: a 5x5 pixel convolution kernel can effectively perceive large areas of red cell regions and obvious horizontal edges, quickly locking the approximate boundary line, like using a wide-angle lens for rapid scanning.

[0100] Second, shallow networks correspond to simple features: recognizing a dominant red cell layer does not require as deep a feature hierarchy as distinguishing the buffer layer, and three layers of convolution are sufficient to extract basic features such as edges -> textures -> object parts to complete the positioning task.

[0101] Finally, small models achieve high real-time performance: during the fastest settling period, the model's processing speed can keep up with the rapid changes in the interface, ensuring real-time monitoring.

[0102] Working principle of the second model (for the white blood cell-platelet formation period)

[0103] 1. Model architecture:

[0104] The second model is designed specifically for the white blood cell-platelet formation period (brown-yellow layer formation period). Its architecture includes, in order: an input layer (128x128 pixel grayscale image), a convolution layer 1 (using 32 5x5 pixel filters), a convolution layer 2 (using 64 5x5 pixel filters), a convolution layer 3 (using 128 3x3 pixel filters), a convolution layer 4 (using 128 3x3 pixel filters), a max pooling layer (2x2 pixel window, stride 2), a fully connected layer 1 (256 neurons), a fully connected layer 2 (128 neurons), and an output layer (3 neurons).

[0105] 2. Model characteristics and benefits:

[0106] With a deeper 4-layer convolutional structure, the network is allowed to build more complex feature hierarchies. From simple edge and color features, to complex texture and patterns, it is eventually able to recognize subtle differences between the top red blood cell layer, the white blood cell layer, the platelet layer, and other buffer layer structures.

[0107] The first two layers use a 5x5 pixel convolutional kernel for preliminary macro positioning, and the last two layers switch to a 3x3 pixel convolutional kernel for fine feature extraction. With a coarse-to-fine strategy, the receptive field and parameter efficiency are balanced.

[0108] The increasing number of filters (32 -> 64 -> 128 -> 128) and larger fully connected layers (256 -> 128 neurons) provide sufficient capacity for the model to learn and distinguish a variety of subtle features.

[0109] 3. Coordination with sedimentation characteristics:

[0110] For example, Figure 3 The formation of the buffer layer (white blood cells-platelets) is shown in the infrared image of blood separation. The core feature of the white blood cell-platelet formation period is that the target is fine (the brown-yellow layer is very thin) and the features are complex (multi-layer structure, weak contrast).

[0111] To this end, the second model uses a deeper network to handle complex features, and the recognition of the brown-yellow layer requires understanding of more abstract feature combinations. A deeper network can combine edge information at the bottom layer into a higher layer texture concept (such as the dense texture at the top of the red blood cell and the white and fluffy texture of the brown-yellow layer).

[0112] Weak and diverse visual features require a larger number of filters to capture in parallel, ensuring that key information is not missed.

[0113] The second model first roughly locates the cell layer area, and then focuses on the interface for fine analysis, which is consistent with the logic of the operator first finding the approximate area and then carefully observing the buffer layer.

[0114] Working principle of the third model (for blood clarification period)

[0115] 1. Model architecture:

[0116] The third model is designed for the blood clarification period. Its architecture includes, in order: input layer (128x128 pixel grayscale image), convolutional layer 1 (using 64 3x3 pixel filters), convolutional layer 2 (using 128 3x3 pixel filters), convolutional layer 3 (using 256 3x3 pixel filters), convolutional layer 4 (using 256 3x3 pixel filters), max pooling layer (2x2 pixel window, stride 2), fully connected layer 1 (512 neurons), fully connected layer 2 (256 neurons), output layer (3 neurons).

[0117] 2. Model characteristics and benefits:

[0118] The third model has the deepest and widest architecture among the three CNN models. With 4 layers of convolution and the largest number of filters (64 -> 128 -> 256 -> 256), combined with fully connected layers (512 -> 256 neurons), it provides the strongest learning ability and feature expression ability.

[0119] All use 3x3 pixel small convolution kernel stacking, increase the network depth to improve the nonlinear expression ability, so as to learn very abstract and high-level features (such as blood uniformity, clarity), while the parameter efficiency is higher than using large convolution kernel.

[0120] This model can go beyond specific edges and textures, and judge the essential characteristics of plasma from a global and overall perspective, such as identifying minor changes in spectral characteristics caused by slight hemolysis or minor turbidity at the interface.

[0121] 3. Relationship with sedimentation characteristics:

[0122] For example Figure 4 As shown in the blood separation infrared image during the plasma clarification period, the core feature of the blood clarification period is feature abstraction (the plasma layer itself has no obvious texture), and the judgment relies on global information and high-level semantics (such as color uniformity, transparency).

[0123] In view of the characteristics of the blood clarification period, the design of the third model:

[0124] Use the deepest and widest network to process the most abstract features. Judging whether the plasma is completely clarified is a high-level semantic task. Similar to experienced experts relying on feelings to judge, this feeling comes from the deep synthesis of a large number of underlying features. The depth and width of the third model enable it to complete this complex feature abstraction process.

[0125] Small convolution kernel stacking realizes the layer-by-layer abstraction process. The network forces the signal to pass through multiple layers of nonlinear transformation, each layer generalizes and combines the features of the previous layer, and finally forms an abstract feature representation related to the ideal clarification state at the top layer.

[0126] For high-dimensional abstract features from deep convolutional layers, a large-capacity decision layer is needed to accurately map to the final interface coordinates and area, ensuring the highest accuracy in the terminal stage.

[0127] Specifically, the training of the convolutional neural network model includes:

[0128] Obtain an infrared image sample set labeled with plasma region boundaries and separation interface positions;

[0129] The three convolutional neural network models are independently trained using a large number of infrared images labeled with corresponding stage features, and the model parameters are optimized by a mean square error loss function. Through large-scale sample training, the generalization ability of the model is enhanced, ensuring stable feature extraction in complex scenes (such as different blood samples and device noise fluctuations), and reducing errors caused by sample differences.

[0130] In one specific embodiment, the present application discloses an infrared monitoring data processing method for blood separation, wherein the method comprises:

[0131] An infrared image acquisition step is configured to capture a sequence of infrared images of the blood separation test tube in the wavelength range of 8-14 microns (μm) by an infrared camera at a rate of 30 frames per second (fps), capturing the thermal radiation distribution characteristics of the blood components;

[0132] A dynamic noise suppression step is configured to use a sliding window Kalman filter (SW-KF) algorithm to perform time domain noise reduction on the image sequence, to calculate the observation noise variance in real time by initializing a sliding window of 50 frames, and to dynamically switch between a high gain mode (SNR≥20 dB) or a low gain mode (SNR<20 dB) according to the signal-to-noise ratio (SNR) to balance the tracking speed and noise suppression;

[0133] A feature enhancement and extraction step is configured to process the noise-reduced images based on a multi-model switching convolutional neural network architecture, the system dynamically selects and loads a convolutional neural network model matching the current blood separation stage according to the centrifugation time or the identified interface features, to extract the blood separation interface coordinates and the plasma area; the convolutional neural network model matching the current blood separation stage is independently trained using more than 1500 groups (500 groups per model) of infrared images with normal and abnormal samples;

[0134] A parameter generation and warning step is configured to calculate key separation parameters, including the plasma layer height (Hplasma) ) and hematocrit (Hct) ), and to trigger a warning signal when or the interface fluctuation standard deviation , and to output the separation interface dynamic curve and real-time parameter values.

[0135] In this embodiment, the specific implementation steps include:

[0136] S1. Infrared image acquisition

[0137] (1) The infrared camera acquires blood separation test tube images at a rate of 30 fps, with a wavelength range of 8-14 µm;

[0138] (2) The image sequence reflects the dynamic process of blood layering.

[0139] S2. Dynamic noise suppression (SW-KF)

[0140] (1) Initialize window (L=50 frames), calculate observation noise variance in real time:

[0141] ;

[0142] (2) Dynamic switching of filtering mode:

[0143] ① High-gain mode (SNR≥20dB): rapid tracking of interface mutation;

[0144] ② Low-gain mode (SNR<20dB): suppress environmental noise interference.

[0145] S3. CNN feature extraction based on multi-model switching

[0146] (1) Model switching mechanism: the system dynamically switches and loads among three convolutional neural network models matched with the current blood separation stage according to the time threshold of centrifugation process or the interface morphological features (such as interface number, clarity) identified in the previous frame:

[0147] In the early stage of separation (such as the first 2 minutes), or when the interface is single and significant, load the first model (for the rapid sedimentation period of red blood cells).

[0148] In the middle stage of separation (such as around 5 minutes), or when multiple closely adjacent interfaces are identified, load the second model (for the white blood cell-platelet formation period).

[0149] In the late stage of separation (such as around 8 minutes), or when the upper area is homogeneous and clear, load the third model (for the blood clarification period).

[0150] (2) Architecture and function of each model matched with the current blood separation stage:

[0151] ① First model (rapid sedimentation period): adopt a shallow and wide initial design, its architecture is: Convolutional layer 1 (32 5x5 pixel filters) → Convolutional layer 2 (64 3x3 pixel filters) → Convolutional layer 3 (64 3x3 pixel filters) → Max pooling layer → Fully connected layer 1 (128 neurons) → Fully connected layer 2 (64 neurons) → Output layer. This model uses a larger initial receptive field (5x5 pixel convolution kernel) to quickly capture the significant and macroscopic interface between the red blood cell layer and the plasma, achieving high-efficiency preliminary positioning.

[0152] ② The second model (buffer layer formation period): a deeper and more refined design is adopted, with the architecture being: Convolutional layer 1 (32 5x5 pixel filters) -> Convolutional layer 2 (64 5x5 pixel filters) -> Convolutional layer 3 (128 3x3 pixel filters) -> Convolutional layer 4 (128 3x3 pixel filters) -> Max pooling layer -> Fully connected layer 1 (256 neurons) -> Fully connected layer 2 (128 neurons) -> Output layer. This model focuses on extracting subtle interface features between the red blood cell layer and the buffy coat layer, and the buffy coat layer and the plasma layer, by using a deeper network and increasing the number of filters layer by layer, to solve the problem of buffer layer recognition.

[0153] ③ The third model (blood clarification period): the deepest and widest design is adopted, with the architecture being: Convolutional layer 1 (64 3x3 pixel filters) -> Convolutional layer 2 (128 3x3 pixel filters) -> Convolutional layer 3 (256 3x3 pixel filters) -> Convolutional layer 4 (256 3x3 pixel filters) -> Max pooling layer -> Fully connected layer 1 (512 neurons) -> Fully connected layer 2 (256 neurons) -> Output layer. This model improves the ability to judge abstract features (such as plasma clarity and final interface clarity) by stacking a large number of small convolutional kernels to build a deep network, ensuring the highest accuracy of end-point monitoring.

[0154] (3) Training and output: all the above models use ReLU activation function and are optimized with mean square error loss function. They receive denoised infrared images and finally output accurate separation interface coordinates and plasma area .

[0155] S4. Parameter generation and early warning

[0156] (1) Calculate key parameters:

[0157] ① Plasma layer height: ;

[0158] wherein represents the pixel area of the plasma region in the infrared image, which refers to the pixel area occupied by the plasma part in the denoised image output by the CNN feature extraction step.

[0159] represents the total pixel area of the cross-section of the test tube, specifically the total pixel area of the cross-section of the blood separation test tube in the infrared image, which is extracted from the test tube contour by image segmentation technology, and the plasma area uses the same pixel reference.

[0160] represents the actual physical height of the test tube (unit: mm), which refers to the real physical height of the standard blood separation test tube used in the experiment, and the conversion logic is represents the area ratio of the plasma layer in the cross-section of the test tube, multiplied by The area ratio can be converted into the actual plasma layer height (unit: mm).

[0161] ②Hematocrit: ;

[0162] In the formula represents the pixel area of the plasma region in the infrared image, and have exactly the same meaning, and is used uniformly in the hematocrit formula. is consistent with in the plasma layer height formula.

[0163] represents the total pixel area of the test tube cross-section, and is exactly the same as in the plasma layer height formula.

[0164] is the area ratio of the plasma layer, which is the area ratio of the blood cell layer (mainly red blood cells). Since the blood cells are dense after centrifugation, the cross-sectional area ratio can directly reflect the volume concentration, so this ratio is the hematocrit.

[0165] (2) Abnormality determination: if or the interface fluctuation standard deviation , an alarm is triggered, and the alarm interface is shown in Figure 5 .

[0166] Example Two

[0167] As shown in Figure 6 , the present embodiment provides a non-invasive blood separation process monitoring system based on infrared imaging, comprising:

[0168] An infrared image acquisition module: for acquiring an infrared image sequence of a blood separation device through an infrared camera, capturing the thermal radiation distribution characteristics of blood components;

[0169] A dynamic noise suppression module: for performing time-domain noise reduction on the infrared image sequence using a sliding window Kalman filter algorithm, dynamically updating the noise covariance and adaptively adjusting the sliding window length, and outputting the noise-reduced infrared image;

[0170] Feature enhancement and extraction module: used for processing the denoised infrared image based on the multi-model switching convolutional neural network architecture, dynamically selecting and loading the convolutional neural network model matched with the current blood separation stage to extract the blood separation interface morphology feature; the convolutional neural network model matched with the current blood separation stage includes a first model optimized for the erythrocyte rapid sedimentation period, a second model optimized for the white blood cell-platelet formation period and a third model optimized for the blood clarification period;

[0171] Parameter generation and early warning module: used for calculating key separation parameters based on the separation interface morphology feature, and triggering a warning signal when the parameters are abnormal.

[0172] Embodiment three

[0173] Embodiment three of the present application provides an electronic device.

[0174] An electronic device includes a memory, a processor, and a program stored on the memory and running on the processor, and the processor implements the steps in the non-invasive blood separation process monitoring method based on infrared imaging as described in embodiment one of the present application.

[0175] The detailed steps are the same as the non-invasive blood separation process monitoring method based on infrared imaging provided in embodiment one, and will not be repeated here.

[0176] Embodiment four

[0177] Embodiment four of the present application provides a computer readable storage medium.

[0178] A computer readable storage medium has a program stored thereon, and the program is executed by a processor to implement the steps in the non-invasive blood separation process monitoring method based on infrared imaging as described in embodiment one of the present application.

[0179] The detailed steps are the same as the non-invasive blood separation process monitoring method based on infrared imaging provided in embodiment one, and will not be repeated here.

[0180] Embodiment five

[0181] Embodiment five of the present application provides a computer program product.

[0182] A computer program product includes software code, and the program in the software code performs the steps in the non-invasive blood separation process monitoring method based on infrared imaging as described in embodiment one of the present application.

[0183] The detailed steps are the same as the non-invasive blood separation process monitoring method based on infrared imaging provided in embodiment one, and will not be repeated here.

[0184] Those skilled in the art will appreciate that embodiments of the application can be devised for a method, a system, or a computer program product. Accordingly, the present application can be embodied in the form of an entirely hardware embodiment, an entirely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present application can take the form of a computer program product on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROMs, optical storage devices, etc.) embodying computer readable program code. Embodiments of the present application can be implemented with various computer program languages such as the object-oriented programming language Java and the interpreted scripting language JavaScript, etc.

[0185] The present application is described in reference to the flowchart illustrations and / or block diagrams according to the embodiments of the application. It is to be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general purpose computer, special purpose computer, embedded processing device or other programmable data processing device to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing device, create means for implementing the functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams.

[0186] These computer program instructions can also be stored in a computer- readable memory that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable memory produce an article of manufacture including instructions which implement the functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams.

[0187] These computer program instructions can also be loaded onto a computer or other programmable data processing device to cause a series of operational steps to be performed on the computer or other programmable device to produce a computer-implemented process such that the instructions which execute on the computer or other programmable device provide steps for implementing the functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams. Figure 1 one or more functions specified in the flowchart illustrations and / or block diagrams.

[0188] The above description is only preferred embodiment of the present application and is not intended to limit the present application. The present application can have various modifications and changes for those skilled in the art. Any modification, equivalent replacement, improvement, etc. made within the spirit and principle of the present application shall be included in the protection scope of the present application.

Claims

1. A non-invasive blood separation process monitoring method based on infrared imaging, characterized in that, The application relates to a blood separation interface state monitoring method based on infrared image processing. An infrared image sequence of a blood separation device is collected by an infrared camera to capture the thermal radiation distribution characteristics of blood components; A sliding window Kalman filtering algorithm is used to perform time domain noise reduction on the infrared image sequence, dynamically update noise covariance and adaptively adjust the length of the sliding window, and output the noise-reduced infrared image; A multi-model switching convolutional neural network architecture is used to process the noise-reduced infrared image, dynamically select and load a convolutional neural network model matched with the current blood separation stage to extract blood separation interface morphological features; The convolutional neural network model matched with the current blood separation stage comprises a first model optimized for a red blood cell rapid sedimentation period, a second model optimized for a white blood cell-platelet formation period and a third model optimized for a blood clarification period; The first model is used for the red blood cell rapid sedimentation period, and the architecture comprises three convolutional layers, filter numbers of which are 32, 64 and 64 respectively, and sizes of which are 5*5 pixels, 3*3 pixels and 3*3 pixels respectively; a maximum pooling layer; and two fully connected layers, neuron numbers of which are 128 and 64 respectively; The second model is used for the white blood cell-platelet formation period, and the architecture comprises four convolutional layers, filter numbers of which are 32, 64, 128 and 128 respectively, and sizes of which are 5*5 pixels, 5*5 pixels, 3*3 pixels and 3*3 pixels respectively; a maximum pooling layer; and two fully connected layers, neuron numbers of which are 256 and 128 respectively; The third model is used for the blood clarification period, and the architecture comprises four convolutional layers, filter numbers of which are 64, 128, 256 and 256 respectively, and sizes of which are all 3*3 pixels; a maximum pooling layer; and two fully connected layers, neuron numbers of which are 512 and 256 respectively; The convolutional neural network model matched with the current blood separation stage adopts a ReLU activation function and outputs a separation interface coordinate and a plasma region pixel area ; Key separation parameters are calculated based on the separation interface morphological features, and a warning signal is triggered when the parameters are abnormal.

2. The method of claim 1, wherein, The dynamic noise covariance updating comprises: A sliding window with a length of L is initialized, and the observation noise variance is calculated in real time: ； wherein, is the i-th frame of original infrared image data, is the Kalman filter prediction value, is the current frame number, is the sliding window length.

3. The method of claim 1, wherein, The time domain noise reduction process comprises dynamically switching filtering modes: When the signal-to-noise ratio is greater than or equal to a first threshold value, a high-gain mode is enabled, and the Kalman gain coefficient is increased to accelerate tracking of separation interface mutations; When the signal-to-noise ratio is less than the first threshold value, a low-gain mode is enabled, and the Kalman gain coefficient is reduced to suppress environmental noise.

4. The method of claim 1, wherein, The key separation parameters include plasma layer height and hematocrit ; The height of the plasma layer The formula for calculating the height of the plasma layer is: ； The hematocrit The formula for calculating the hematocrit is: ； wherein, is the area of the plasma region pixels, is the total area of the container cross section pixels, is the physical height of the container.

5. The method of claim 4, wherein, The parameter abnormality triggering condition is: or the standard deviation of the isolated interface fluctuation ; wherein, is a hematocrit warning threshold, is an interface stability threshold.

6. An infrared imaging-based non-invasive blood separation process monitoring system, characterized by The application relates to a blood separation interface state monitoring method based on infrared image processing. An infrared image sequence of a blood separation device is collected by an infrared camera to capture the thermal radiation distribution characteristics of blood components; A sliding window Kalman filtering algorithm is used to perform time domain noise reduction on the infrared image sequence, dynamically update noise covariance and adaptively adjust the length of the sliding window, and output the noise-reduced infrared image; A multi-model switching convolutional neural network architecture is used to process the noise-reduced infrared image, dynamically select and load a convolutional neural network model matched with the current blood separation stage to extract blood separation interface morphological features; The convolutional neural network model matched with the current blood separation stage includes a first model optimized for the erythrocyte fast settling period, a second model optimized for the leukocyte-platelet formation period, and a third model optimized for the blood clarification period; The first model is used for the erythrocyte fast settling period, and the architecture thereof includes three convolutional layers, the filter numbers of which are 32, 64, and 64 respectively, and the sizes of which are 5*5 pixels, 3*3 pixels, and 3*3 pixels respectively; a max pooling layer; and two fully connected layers, the neuron numbers of which are 128 and 64 respectively; The second model is used for the leukocyte-platelet formation period, and the architecture thereof includes four convolutional layers, the filter numbers of which are 32, 64, 128, and 128 respectively, and the sizes of which are 5*5 pixels, 5*5 pixels, 3*3 pixels, and 3*3 pixels respectively; a max pooling layer; and two fully connected layers, the neuron numbers of which are 256 and 128 respectively; The third model is used for the blood clarification period, and the architecture thereof includes four convolutional layers, the filter numbers of which are 64, 128, 256, and 256 respectively, and the size of which is 3*3 pixels; a max pooling layer; and two fully connected layers, the neuron numbers of which are 512 and 256 respectively; The convolutional neural network model matched with the current blood separation stage adopts a ReLU activation function and outputs a separation interface coordinate and a plasma region pixel area ; The parameter generation and early warning module is used for calculating key separation parameters based on the separation interface morphological features, and triggering an early warning signal when the parameters are abnormal.

7. An infrared imaging based non-invasive blood separation process monitoring device, characterized in that, The device includes a memory and a processor; the memory is used for storing a computer program; the processor is used for implementing the non-invasive blood separation process monitoring method based on infrared imaging as claimed in any one of claims 1 to 5 when executing the computer program.

8. A computer-readable storage medium, characterized in that, The storage medium has a computer program stored thereon, and the computer program, when executed by a processor, implements the non-invasive blood separation process monitoring method based on infrared imaging as claimed in any one of claims 1 to 5.

9. A computer program product comprising software code, characterized in that, The program in the software code performs the steps of the non-invasive blood separation process monitoring method based on infrared imaging as claimed in any one of claims 1 to 5.

Images