Smart intraoperative bleeding volume image assessment system

Abstract

The present application relates to the technical field of image processing and intelligent evaluation system, and discloses an intelligent intraoperative bleeding amount image evaluation system, comprising: a target classification unit, used for distinguishing pooled blood and surgical dressings based on texture features and contour features; a target tracking unit, used for tracking the surgical dressing and predicting an expected position area thereof; and an evaluation logic control unit, used for forcing the target falling into the expected position area to inherit the identity of the surgical dressing, and bypassing the re-judgment of the target classification unit based on texture, wherein the present application constructs an identity inheritance mechanism based on time domain information, uses the position prediction result of target tracking to preferentially confirm the target identity, and denies the classification result based on the current frame spatial features, thereby solving the tracking interruption problem caused by the loss of visual features due to dressing infiltration.

Description

TECHNICAL FIELD

[0001] The present application relates to an intelligent intraoperative bleeding volume image evaluation system, belonging to the technical field of image processing and intelligent evaluation system. BACKGROUND

[0002] Current image analysis techniques are used as objective quantitative evaluation means in many fields to establish the corresponding relationship between the color, texture or morphological features of the image and the specific physical quantity by extracting the features. However, the effectiveness of such technical methods is usually based on a key implicit premise that the key visual features of the measured target and the environmental conditions of the collected image can remain relatively stable and consistent. When conventional image processing methods are applied to complex scenes that are not controlled and dynamically changing, this premise no longer holds. For example, in a surgical environment, the movement of light sources, the reflection of instruments, and the intervention of flushing liquids all cause instantaneous or continuous interference with the original visual features of the image. This interference is not simply noise, but fundamentally changes the original morphology of the target features in the image, and even mixes with the target itself.

[0003] Under such working conditions, merely trying to improve the robustness of the static evaluation model or filtering image noise is difficult to address the root of the problem. The stability of the visual features such as color and texture that the system relies on has been destroyed, and the image data that the system faces not only reflects the true target state, but also mixes in artifacts introduced by dynamic changes in the environment. The evaluation system lacks an internal mechanism to actively identify the extent to which the current image data is affected by such dynamic interference. The target such as the dressing and the pooled blood has been visually confused, and even the physical state of the target such as blood itself has fundamentally changed from flowing to coagulation. In order to solve the limitations of visual evaluation and conventional image processing, the technical solution shifts to a quantitative method based on hardware sensors. For example, the Chinese invention patent with publication number CN118236043A discloses a bleeding volume quantification monitoring method and system based on an intraoperative bleeding volume quantification device. This solution attempts to collect bleeding from different sources through hardware quantification devices such as a self-vaginal blood collection assembly, a negative pressure suction bag, and an autologous blood recovery machine, and to calculate the bleeding volume based on flow or weight data. However, this method highly depends on the complex and invasive physical collection pipeline, and its premise is that all bleeding can be effectively captured and introduced into the corresponding sensor, which leads to its inability to evaluate the pooled blood that has not been collected and remains in the surgical field. Moreover, the complex hardware deployment also faces limitations in different surgical scenarios.

[0004] Therefore, how to design an image evaluation system which can not only analyze the target features, but also actively perceive and quantify the dynamic interference and the state change of the target in the environment, and dynamically adaptively regulate the evaluation model and processing logic according to the perception results, so as to obtain reliable evaluation results in the non-controlled complex working conditions, has become a technical problem to be solved by the present application. SUMMARY

[0005] The present application provides an intelligent intraoperative bleeding amount image evaluation system, which mainly aims to solve the problem that the existing image evaluation system lacks real-time perception and self-regulation ability for interference and state change when facing dynamic and high-interference complex environment, resulting in evaluation logic failure and result distortion.

[0006] To achieve the above-mentioned purpose, the present application provides an intelligent intraoperative bleeding amount image evaluation system, which comprises:

[0007] An image feature extraction unit is configured to extract texture features and contour features in the obtained surgical region image.

[0008] A target classification unit is configured to identify a target region in the surgical region image based on the texture features and the contour features, and distinguish the target region as pooled blood or surgical dressing.

[0009] A target tracking unit is configured to start target tracking on the target region identified as surgical dressing by the target classification unit, and predict an expected position region of the target region in a subsequent image frame.

[0010] An evaluation logic control unit is configured to determine the target region falling within the expected position region as surgical dressing, and bypass the re-determination of the target region by the target classification unit based on the texture features when the target region is determined as surgical dressing due to falling within the expected position region.

[0011] Preferably, the system further comprises a bleeding amount evaluation unit configured to calculate the blood amount of the target region distinguished as pooled blood by the target classification unit and the target region distinguished as surgical dressing by the target classification unit based on the image features of the target region; the blood amount calculation comprises: analyzing the red saturation and area of the target region, and mapping the red saturation and area to a preset volume model to obtain the blood amount attributed to the pooled blood and the blood amount absorbed by the surgical dressing, respectively.

[0012] Preferably, the bleeding amount evaluation unit comprises an evaluation model; the system further comprises an illumination calibration unit, which is configured to extract image statistics features representing illumination stability of the surgical region images in parallel when evaluating the bleeding amount; the illumination calibration unit is further configured to determine the credibility of the surgical region images for bleeding amount evaluation based on the image statistics features, and to perform dynamic weighting processing on a sequence of bleeding amount evaluation results output by the evaluation model within a time window according to the credibility.

[0013] Preferably, the illumination calibration unit further comprises a baseline reset module; the baseline reset module is configured to perform long-term analysis on a time sequence of the image statistics features; the baseline reset module is further configured to determine that the visual acquisition baseline has been systematically deviated when it is determined that the values of the image statistics features continuously exceed a preset offset threshold; and the baseline reset module is further configured to trigger recalibration of baseline parameters of the evaluation model in response to the systematic deviation.

[0014] Preferably, the system further comprises a dilution compensation unit; the dilution compensation unit comprises a preset dilution compensation evaluation model; the dilution compensation unit is configured to extract second visual features representing the presence of the irrigation liquid in the surgical region images in parallel, the second visual features being high light mirror reflection features and transient deformation features; the dilution compensation unit is further configured to generate a dynamic dilution mask covering the area affected by the irrigation liquid based on the second visual features; and the dilution compensation unit is further configured to automatically invoke the dilution compensation evaluation model to calculate the bleeding amount for the area covered by the dynamic dilution mask.

[0015] Preferably, the system further comprises a dynamic adaptation unit; the dynamic adaptation unit comprises a set of preset bleeding modes; the dynamic adaptation unit is configured to extract dynamic process features representing the morphological expansion speed or color change rate of the target region based on a continuous sequence of the surgical region images; the dynamic adaptation unit is further configured to classify the current bleeding process into one of the preset bleeding modes according to the dynamic process features; and the dynamic adaptation unit is further configured to adaptively adjust the internal parameters of the evaluation model based on the bleeding mode.

[0016] Preferably, the dynamic adaptation unit extracts the dynamic process features representing the color change rate by calculating the Bhattacharyya distance between the color histograms of the target region and in the consecutive image frames and , which are distinguished as pooled blood or surgical dressings. The calculation of the Bhattacharyya distance satisfies: wherein and are the color histograms of the target region in the consecutive image frames and . a group index of the histogram, a total number of groups, a natural logarithm function.

[0017] Preferably, the dynamic self-adaptive unit is further configured to acquire intermediate information generated in the calculation of the dynamic process feature when extracting the dynamic process feature, the intermediate information representing a pixel region with a calculation failure or a residual error that is too large; the dynamic self-adaptive unit is further configured to identify a microstate feature of a target region classified as pooled blood based on the intermediate information, the microstate feature being a coagulation state, and the dynamic self-adaptive unit is further configured to correct an evaluation logic of the evaluation model using the microstate feature, the corrected evaluation logic excluding the region identified as the coagulation state from the calculation of the active bleeding amount.

[0018] Preferably, the image statistical feature includes a histogram distribution feature of the image, the histogram distribution feature being used to quantify a proportion of high-light region pixels and a proportion of dark pixels, and a high-frequency information feature of the image, the high-frequency information feature being used to quantify a strong edge energy caused by specular reflection.

[0019] Preferably, the evaluation logic control unit is further configured to count a number of targets currently located in the image of the surgical region and identified as the surgical dressing by the target classification unit and tracked by the target tracking unit in real time, and compare the number of targets with a preset reference number of dressings.

[0020] Compared with the prior art, the present application has the following beneficial effects:

[0021] 1. The present application provides a parallel analysis and dynamic regulation mechanism, the system extracts statistical features representing the current image acquisition quality while performing image evaluation; the system uses the statistical features to determine the effectiveness of the current data frame for evaluation, and dynamically regulates the evaluation results in the time sequence based on the effectiveness, which makes the final output of the system based on a series of image data recognized as effective by the system, avoiding the distortion of the evaluation results caused by instantaneous image acquisition artifacts.

[0022] 2. The image evaluation is divided into two levels that work cooperatively, i.e., dynamic process features are extracted using the time domain information of the image sequence, and the current working condition is classified into a preset mode based on the process features, the classification result of the mode is used as a regulation signal to feed back to the main evaluation model to adaptively adjust the internal parameters of the main evaluation model, which makes the evaluation model able to match the working condition in real time, improving the evaluation continuity of the system when switching between different working conditions.

[0023] 3、The application reuses the front-end data of image analysis, extracts texture and contour features of the same area in parallel while analyzing color features; the system realizes the classification of the target through the combination of these composite features, that is, to distinguish the targets with similar visual features but different physical natures, and the classification result is further used to logically attribute the evaluated total amount to different physical carriers, so that the output of the system is changed from a single total amount to a structured and traceable classification data. BRIEF DESCRIPTION OF DRAWINGS

[0024] Fig. 1 It is a schematic diagram of data flow and dynamic regulation loop of the system function module of the application;

[0025] Fig. 2 It is a comparison diagram of the correction effect of the condensation state recognition on the evaluation of active bleeding amount;

[0026] Fig. 3 It is a hardware deployment architecture and core processing node function division diagram of the system of the application. DETAILED DESCRIPTION

[0027] In order to make the personnel in the technical field better understand the technical solutions in the application, the technical solutions in the embodiments of the application will be clearly and completely described below, obviously, the described embodiments are only part of the embodiments of the application, not all the embodiments, based on the embodiments in the application, all other embodiments obtained by the person skilled in the art without creative labor should belong to the protection scope of the application.

[0028] The embodiment of the present application provides a kind of intelligent intraoperative bleeding amount image evaluation system, this system can be deployed in hardware in the surgical auxiliary system integrated with image acquisition device or independent computing server, it is used to realize the objective evaluation of bleeding amount in dynamic, high interference environment in surgical procedure by a series of visual intelligent analysis and logic control unit;System can be divided into image feature extraction unit, target classification unit, target tracking unit and evaluation logic control unit in function;Image feature extraction unit, for extracting the texture feature and contour feature required for subsequent analysis from the continuous acquisition surgical area image frame;Target classification unit, then qualitative identification is carried out to the target in image according to texture feature and contour feature;Target tracking unit is used to maintain the continuous cognition of specific target in time sequence, and evaluation logic control unit is then according to specific rule coordination analysis result of the foregoing unit, to output final evaluation data;In one specific embodiment, the input processed by image feature extraction unit is the original pixel stream from surgical area image acquisition device, and its task is to extract texture feature and contour feature for the region in image in parallel;In engineering implementation, the extraction of texture feature can use local binary pattern (LBP) algorithm, which generates a histogram representing local texture by comparing the gray level relationship between center pixel and its neighborhood pixel, which can be used to describe the fabric grid structure specific to surgical dressing;At the same time, the extraction of contour feature can use Canny edge detection algorithm, which can demarcate the outer boundary of target through multi-stage processing including Gaussian filtering, gradient calculation, non-maximum suppression and double threshold detection, and the output edge atlas is used to represent the morphological integrity of target;In this way, the final output of this unit is a composite feature vector, which contains the quantitative description of texture and morphology of each target in surgical area, and is submitted to target classification unit for subsequent processing;After receiving the above composite feature vector, target classification unit is used to solve the confusion problem of visual features similar but physically different targets in surgical field, specifically, that is, to distinguish between pooled blood and surgical dressing.

[0029] To achieve this classification, the unit incorporates a decision logic based on a combination analysis of the input feature vectors: when a target region is determined to have low-frequency fluid texture (i.e., a flat LBP histogram distribution and irregular contours, i.e., diffuse Canny edge map morphology), the target is classified as pooled blood; conversely, when a target region is determined to have high-frequency fabric texture (i.e., an LBP histogram exhibiting specific peaks and relatively regular contours), the target is classified as surgical dressing. This classification result is used, on the one hand, to structurally attribute the subsequently calculated blood volume, and on the other hand, to trigger the activation of the target tracking unit; the target tracking unit is specifically activated for targets identified as surgical dressings by the target classification unit. In visual tracking, a problem exists that targets such as surgical dressings, after absorbing blood, gradually lose the visual features used for initial identification, such as fabric texture; to address this, the system employs an identity inheritance mechanism based on temporal information; specifically, when a target in... Once a time point is identified as surgical dressing by the target classification unit, the target tracking unit, such as a Kalman filter or a kernel correlation filter tracker, immediately starts tracking it based on its current time point. The state of the image is used to predict its position in the next image frame. The expected location region at time; this expected location region is a region of interest (ROI) that includes a certain tolerance margin in space; the evaluation logic control unit is responsible for executing the above identity inheritance mechanism, and its operating logic is as follows: in At that moment, the system detected a new target falling into the area tracked by the target tracking unit. Within the predicted location area; at this time, the evaluation logic control unit forcibly identifies the target falling into the area as a surgical dressing, allowing it to directly inherit... The identity of the target at any given time; and when this unit performs this determination, the bypass target classification unit makes a re-determination of the target region based on texture features. This mechanism uses the position prediction results of target tracking to prioritize the confirmation of the target identity, and maintains continuous recognition and tracking of its identity even when the target's visual features have become invalid.

[0030] After classifying and tracking the targets, the system also includes a bleeding volume assessment unit. This unit calculates the blood volume for target areas classified as pooled blood and target areas classified as surgical dressings, respectively. The calculation process is also based on visual feature analysis, specifically analyzing the red saturation and area of ​​the target area. Taking the HSV color space (which includes hue, saturation, and lightness components) as an example, the system calculates the average saturation value of all pixels within the target area. And the total number of pixels, i.e., the area ;Will and These two parameters are input together into a preset volume model. Mapping is performed in the volume model; This can be a function or lookup table established through offline calibration, for example, by soaking a standard dressing with different volumes of blood under controlled light and recording the corresponding values. and The system calculates the value to fit the mapping relationship; thus, it can separately obtain the blood volume attributed to the collected blood and the blood volume attributed to the surgical dressing, achieving a structured classification and attribution of bleeding volume; the lighting conditions in the surgical environment may be unstable, such as the movement of the surgical light or instrument reflection, which will directly affect the red saturation used for assessment. Visual features such as bleed can lead to distorted evaluation results. Therefore, the system of this invention also includes an illumination calibration unit, which quantifies the reliability of the current image's visual features. This illumination calibration unit operates in parallel with the bleed assessment unit, simultaneously extracting image statistical features characterizing illumination stability while assessing bleed. In a specific implementation, the image statistical features may include: the image's histogram distribution features, specifically the proportion of pixels in highlight areas and the proportion of pixels in shadow areas; and the image's high-frequency information features, specifically the strong edge energy detected by the Laplacian operator or the Sobel operator, which is related to the intensity of specular reflection. Based on these statistical features, such as when the proportion of highlight pixels exceeds 20% or the strong edge energy exceeds a preset threshold, the unit determines the reliability of the current image frame for bleed assessment. Finally, the unit determines the confidence level based on this confidence level. The assessment model outputs a sequence of bleeding assessment results within a time window (e.g., the past 5 seconds). Dynamic weighting can be performed, and the weighting method can be as follows: ,in, This is the final hemorrhage assessment value after light calibration for the current time window. For the first time window The original hemorrhage assessment value corresponding to the frame image. For the first The credibility value corresponding to the frame image. This indicates that the calculation results of all frames within the time window are summed. In this way, image frames with stable lighting have a high weight in the final result, while the weight of image frames affected by instantaneous reflection interference is automatically reduced by the system.

[0031] Furthermore, while the aforementioned illumination calibration mechanism addresses transient interference, it may continuously output low confidence levels if the visual acquisition baseline experiences a systematic shift, such as when the main light source in the operating room is switched. To address this, the illumination calibration unit also includes a baseline reset module. This module performs long-term analysis on the time series of image statistical features. Its logic is as follows: when the value of an image statistical feature, such as the percentage of pixels in the highlight area, continuously exceeds a preset offset threshold (e.g., the percentage remains greater than 15% for more than 10 seconds), the system no longer considers it transient interference but determines that the visual acquisition baseline has undergone a systematic shift. In response to this determination, the module immediately triggers the evaluation model, i.e. The model's baseline parameters are recalibrated. The system will use the current, new, and stable lighting environment as a benchmark to recalibrate its color-volume mapping relationship, enabling the system to readjust after environmental changes. Furthermore, another source of visual disturbance during surgery is the introduction of irrigation fluid; its dilution effect due to mixing with blood can affect... The system of this invention also includes a dilution compensation unit to achieve the perception of flushing events. This unit extracts second visual features in parallel to characterize the presence of flushing fluid, specifically the specular reflection features and transient deformation features caused by the flow of transparent fluid in the image. Based on these second visual features, the system generates a dynamic dilution mask covering the area affected by the flushing fluid in real time. The system performs spatialized model switching: for areas not covered by the mask, the standard evaluation model A is used; while for areas covered by the mask, i.e., the dilution area, the system automatically calls a preset dilution compensation evaluation model B to calculate the bleeding volume. Model B is specifically calibrated offline for the dilution features of light red, thereby correcting the evaluation distortion caused by the dilution effect.

[0032] The system of this invention also possesses dynamic adaptive capabilities to address the impact of different bleeding patterns, such as slow oozing or rapid gushing, on the static evaluation model. The system includes a dynamic adaptive unit that extracts dynamic process features characterizing the bleeding process rather than static features based on a continuous sequence of surgical area images. These features may include morphological expansion velocity, which can be calculated using optical flow methods to determine the motion vector at the edge of the blood pool or the rate of color change. The quantification of the rate of color change is achieved by calculating the rate of color change across consecutive image frames. and Color histograms of target areas categorized as blood collection or surgical dressings and The distance between the two Bachs This is achieved through the Bach distance. The calculation satisfies: ,in, For the group index of the histogram, The total number of groups It is the natural logarithm function; a higher The value, i.e., the feature distance, indicates a large change in color; based on these dynamic process characteristics, such as high optical flow velocity and large Bartholomew's distance, the dynamic adaptive unit classifies the current bleeding process into a preset bleeding pattern, such as rapid gushing; and based on the result of this pattern classification, it adaptively adjusts the evaluation model (i.e., The model's internal parameters, such as adjusting The sensitivity threshold enables the evaluation model to match the current bleeding condition in real time. In one implementation, the dynamic adaptive unit can also reuse intermediate computational information. Specifically, when extracting dynamic process features using optical flow, the unit obtains the intermediate information generated by its calculation, which characterizes pixel areas where the calculation fails or the residual is too large. In visual analysis, optical flow will fail to calculate in areas where the texture changes abruptly, such as areas where blood coagulates. The unit uses this failure information to perform spatial collaborative determination with color features, such as dark red, thereby identifying the microscopic state characteristics of the bleeding area, i.e., the coagulation state. Finally, the system utilizes this coagulation state characteristic to correct the evaluation logic of the evaluation model. The corrected evaluation logic removes areas identified as coagulated from the calculation of active bleeding, thereby improving the accuracy of the evaluation. Finally, the evaluation logic control unit can also reuse the results of target tracking to realize surgical safety assistance functions. This unit counts in real time the number of targets currently located in the surgical area image that are identified as surgical dressings and are being continuously tracked by the target tracking unit; and compares this real-time visual count value with a preset dressing baseline number entered before surgery, providing objective visual evidence for the counting of surgical instruments and dressings.

[0033] Example 1: In a surgical scenario, At that moment, a surgical dressing was placed in the surgical area; In the image frame at a given time, the image feature extraction unit extracts features, and the target classification unit classifies the target as surgical dressing based on its high-frequency fabric texture features and relatively regular contour features; this classification result triggers the target tracking unit to start tracking the target. At any given time, the target tracking unit is based on The state at a given moment can be used to predict the target's position. The expected location region in the frame; to In a sequence of consecutive image frames, the surgical dressing begins to absorb blood, and the red saturation on its surface increases. As it continues to rise, the high-frequency fabric texture features represented by its LBP texture histogram are gradually covered by the low-frequency fluid texture of blood.

[0034] exist At that moment, the target had become saturated due to blood infiltration, and its original high-frequency fabric texture characteristics were lost, resulting in its current... The spatial visual features of the frame, namely low-frequency fluid texture and irregular contours, converge with the features of the aggregated blood; at this point, the target classification unit is based solely on... The spatial features of the frame are used to determine the target, classifying it as a blood collection. The evaluation logic control unit confirms that the target's spatial location falls within the expected location area predicted by the target tracking unit. Based on this time-domain prediction result, the evaluation logic control unit determines that the target inherits its surgical dressing identity. Furthermore, the evaluation logic control unit bypasses the target classification unit based on the current... The frame texture features are used to re-determine the target; this identity inheritance mechanism is a collaborative operation between the spatial feature analysis of the target classification unit and the temporal position prediction of the target tracking unit, used to deal with situations where the target identity is separated from its instantaneous visual features; in At that moment and in subsequent frames, the system still recognizes the target as a surgical dressing; accordingly, the bleeding assessment unit can continuously call the preset volume model belonging to the surgical dressing. The red saturation and area of ​​the target were calculated, and the calculated blood volume was attributed to the surgical dressing category rather than included in the pooled blood category, thus maintaining the structured classification of bleeding volume assessment even when the visual features of the target were lost.

[0035] Example 2: This example tests the operation of the intelligent intraoperative blood loss image assessment system under two specific conditions: condition one is loss of target visual features, and condition two is sudden environmental interference during data acquisition. The test platform is an external test bench, which includes: an industrial camera with a resolution of 1920x1080 and a frame rate of 30fps for acquiring images of the surgical area; a surgical phantom; a tubing controlled by a micro-injection pump for injecting simulated blood (collected blood) onto the surface of the phantom at a controllable flow rate; a six-axis robotic arm for grasping surgical dressings and placing them in a designated position; and an independent controllable power LED point light source for simulating specular reflection. The system evaluation results are as follows. With a truth value The true value was obtained through the integration of the injection pump flow rate and weighing of the dressing before and after application (weighing accuracy 0.01g). The experiment included control group A, control group C, and the sample group of this invention. Control group A's system only included an image feature extraction unit and a target classification unit, lacking a target tracking unit and an evaluation logic control unit. Control group C's system included target classification and tracking but lacked an illumination calibration unit. The sample group of this invention used the complete system as described in the specific embodiment, including all units. In the target feature annihilation test, at... At that moment, the robotic arm placed a dry surgical dressing into its field of view, its initial... It is 0 mL; At that time, the micro-infusion pump began to drip simulated blood onto the dressing at a rate of 1.0 mL / s, and the dressing began to soak in; At that time, the dressing reached visual complete saturation, and its high-frequency fabric texture characteristics were replaced by low-frequency fluid texture. When the infusion is stopped, the true value of the dressing absorption is... During the experiment, the true value of the collected blood was always 0 mL. The classification and attribution results of the surgical dressing blood volume and collected blood volume of the control group A and the sample group of the present invention were recorded, as shown in Table 1.

[0036] Table 1: Comparison of blood volume attribution under characteristic annihilation conditions;

[0037]

[0038] Table 1 shows that, When the texture features of the dressing are lost due to saturation, the static classifier of control group A identifies the dressing as pooled blood, causing its attribution value for the blood volume of the dressing to drop to 1.9 mL, far below the true value of 11.0 mL, and simultaneously generating an assessment value of 8.5 mL in the pooled blood category; in the sample group of this invention, its evaluation logic control unit activates the identity inheritance mechanism, using the expected location region of the target tracking unit to bypass the current decision of the classification unit, thus in The target was still identified as surgical dressing, and the attribution result of 10.7 mL remained consistent with the true value of 11.0 mL at the end of the trial. An attribution result of 14.8 mL was obtained; in the environmental interference test, At that time, the injection pump began to inject simulated blood into the surface of the phantom at a rate of 0.5 mL / s, forming pooled blood, while the light exposure remained stable; At that time, the LED point light source is turned on to create a high-exposure area of ​​about 40% of the area where blood is collected, i.e., specular reflection; When this happens, the LED point light source is turned off to eliminate interference; Stop the injection at that time; record the assessment values ​​of total bleeding volume for the control group C without light calibration and the sample group of the present invention with light calibration, see Table 2.

[0039] Table 2: Comparison of evaluation results under high light interference conditions;

[0040]

[0041] Table 2 shows that, At that time, highlight interference caused 40% of the blood pool area in the image to be overexposed. The system in control group C classified the highlight area as a non-bleed area, causing its evaluation value to drop from [previous value]. The concentration decreased from 4.4 mL to 2.8 mL, deviating from the true value of 5.5 mL; the light calibration unit of the sample group of this invention... At that time, interference is detected by using image statistical features, namely the proportion of highlight pixels, and the confidence level of the current frame is adjusted accordingly. The concentration dropped to 0.2, and after dynamic weighting, its assessed value of 4.6 mL did not show a significant decrease, consistent with... The assessed value at that time was 4.5 mL. The assessed value of 6.8 mL remained consistent with the trend.

[0042] Example 3: This example combines Figs. 1 to 3 A description of the intelligent intraoperative blood loss image assessment system, such as... Fig. 1 As shown, the system acquires the original image from consecutive image frames of the surgical area. The image feature extraction unit extracts its texture and contour features and outputs a feature vector to the target classification unit. The target classification unit distinguishes between pooled blood and surgical dressing based on the feature vector. The classification result is used to trigger the tracking of surgical dressing and is also submitted to the evaluation logic control unit. The triggered target tracking unit tracks the surgical dressing and predicts its expected location area. This expected location area information is also submitted to the evaluation logic control unit. The evaluation logic control unit executes a key mechanism, namely, identity inheritance and re-determination of bypassed texture features, and submits the confirmed target to the bleeding volume assessment unit. The bleeding volume assessment unit calculates the classified bleeding volume based on the target's red saturation and area, and finally outputs a structured bleeding volume assessment result, which is divided into dressing blood volume / pooled blood volume. At the same time, the system runs two control units in parallel. The illumination calibration unit extracts image statistical features from the time sequence of the original image, determines the assessment confidence, and performs dynamic weighting processing on the bleeding volume assessment unit. The dynamic adaptive unit extracts dynamic process features from consecutive image frames of the surgical area, classifies bleeding patterns, and adaptively adjusts the model parameters of the bleeding volume assessment unit.

[0043] like Fig. 2 As shown, the left vertical axis of the graph represents the assessment value of active bleeding in milliliters (mL), the right vertical axis represents the percentage of the coagulation area in percentages (%), and the horizontal axis represents the time point in seconds (s). The graph displays three curves: a dashed line representing the assessment value of non-coagulation identification, a solid line representing the assessment value of the system of this invention, and a dotted line representing the percentage of the coagulation area. The data shows that as the percentage of the coagulation area increases, the growth rate of the assessment value of the system of this invention, because it excludes the calculation of the coagulation area, is lower than that of the uncorrected non-coagulation identification assessment value, thus achieving a more accurate assessment of active bleeding. Fig. 3As shown, in this architecture, image acquisition devices, such as high-definition cameras in the operating room, are responsible for acquiring image frames of the surgical area and transmitting them to the core processing node. The core processing node can be deployed on an independent computing server or in a surgical assistance system. Internally, it is divided into a main analysis engine and an evaluation and calibration module. The main analysis engine includes an evaluation logic control unit, a target classification unit, a target tracking unit, and an image feature extraction unit. The evaluation and calibration module includes a bleeding assessment unit, an illumination calibration unit, a dilution compensation unit, and a dynamic adaptive unit. The analysis data generated by the main analysis engine is sent to the evaluation and calibration module. This module feeds back the evaluation logic to the main analysis engine and also sends calibration signals to the main analysis engine. The core processing node interacts with an offline calibration model library through model calls and parameter queries. This model library stores preset volume models, dilution compensation models, and bleeding patterns. Finally, the structured evaluation results output by the core processing node are sent to an intraoperative display terminal, such as the screen of the surgical assistance system, for display.

[0044] Example 4: This example illustrates a standardized offline calibration procedure for determining the key internal model and judgment threshold of an intelligent intraoperative blood loss image assessment system, as well as the specific algorithm path for the dynamic adaptive unit to identify the coagulation state. This calibration procedure is executed in a controlled visual experimental environment. The initial state of this environment is defined as follows: using image acquisition equipment consistent with the actual system deployment, with consistent camera model, lens, and working distance; the reference color temperature of the light source is set to 5500K, and the illuminance is stable at 1000 lux. The calibration process steps are as follows: First, calibrate the dilution compensation assessment model; the task involves preparing a set of simulated blood diluents with known concentration gradients (i.e., true values), ranging from 1% to 100%, with a step size of 1%; each concentration of diluent is added to a standard white background plate at a known volume (i.e., true value), and images are acquired simultaneously; for each acquired image, the average red saturation of its diluted region is extracted. With area The calibration result is: to establish a multidimensional lookup table or fit a polynomial function, which will input ( , The first step is to map the data to the corresponding true value of the blood volume, which is the dilution compensation assessment model. The second step is to calibrate the reliability mapping relationship of the illumination calibration unit. The task is as follows: under the reference illumination, acquire a standard red patch image, run the standard assessment model, i.e., model A, and record its assessment value. A controllable power interference light source is introduced to simulate highlight reflection, gradually increasing the proportion of overexposed highlight area in the image from 0% to 50%. At each interference level, image statistical features, i.e., the proportion of highlight area, are extracted in parallel. With strong edge energy And record the current evaluation value. The calibration result is: Calculation of evaluation accuracy. and establish ( , )and The mapping relationship, The value is defined as the confidence level; a specific calibration result is that when At that time, the result obtained from the table lookup The confidence level is 0.15. The third step is to calibrate the offset threshold of the baseline reset module. The task involves continuously acquiring 1000 frames of static background images under a 5500K reference illumination, calculating the statistical features of each frame, using the global histogram entropy as an example. For example, calculate the average value of these 1000 frames of data. and standard deviation The calibration result is: the preset offset threshold is set to a value greater than the normal fluctuation range, and a usable setting value is... When the system detects If the value remains above this threshold for more than 5 seconds, baseline parameter recalibration is triggered. The fourth step is to calibrate the bleeding pattern classification threshold of the dynamic adaptive unit. The task involves simulating two typical operating conditions using an injection pump and acquiring video streams: Condition A is slow oozing at a flow rate of 0.1 mL / s, and Condition B is rapid gushing at a flow rate of 2.0 mL / s. The dynamic process characteristics of both video streams are calculated, specifically the average optical flow vector amplitude at the edge of the bleeding area. Bach distance from color histogram Statistical analysis revealed that operating condition A... Concentrated between 0.5-1.5 pixels / frame, Condition B. Concentrated between 8.0-12.0 pixels / frame; calibration result: set The classification threshold is 5.0 pixels / frame. When the system detects... At that time, the current bleeding pattern is classified as rapid outflow, and the internal parameters of the evaluation model are automatically adjusted.

[0045] The specific operation flow of the algorithm path for identifying the condensation state using intermediate information in the dynamic adaptive unit is as follows: The input is the current image frame. Previous Frame And the binary mask of the dark red area output by the evaluation model. The processing steps include: First, executing a standard optical flow algorithm, such as the Lucas-Kanade method, to calculate... arrive The motion vector field is obtained, and the set of pixels with excessive residuals or tracking failures during the calculation process is acquired to generate a calculation failure mask. The second step is to calculate. and Pixel-level operations are performed to obtain candidate condensation masks. The third step is to... Perform a morphological opening operation, such as erosion and dilation using a 3x3 kernel, to remove isolated noise pixels; in the fourth step, calculate the area of ​​each connected component in the processed mask and discard small connected components with an area less than 50 pixels; the output is the final binarized condensed state mask. ;Should The mask is fed back to the bleeding assessment unit, which will exclude the masked individuals when accumulating active bleeding. The covered pixel area is defined by the rate of color change in the dynamic process characteristics, i.e., the Bach distance. The calibration was also performed using the video streams from step four of Example 4, condition A (slow seepage) and condition B (rapid gushing). Statistical analysis revealed that in condition A, due to the stable blood color, its continuous frames... and of The values ​​are concentrated in the low range of 0.01 to 0.05; while in condition B, due to the rapid coverage of the original area by fresh blood, its The value jumped to the high range of 0.3 to 0.5; based on this data distribution, the calibration result is: set The classification threshold is 0.2. When the system detects... In addition, it can also serve as an auxiliary basis for determining the rapid outflow pattern and trigger adjustments to the internal parameters of the evaluation model.

[0046] Example 5: This example illustrates the offline calibration procedure for constructing a preset volume model, which forms the basis for the bleeding volume assessment unit to calculate the collected blood volume and the blood absorbed by the surgical dressing. The calibration is performed in a controlled visual environment, with the initial state defined as: using image acquisition equipment consistent with the actual system deployment and a 5500K reference light source, with an illuminance stable at 1000 lux. The calibration process includes two parallel stages: In the first stage, for the collected blood model, a micro-injection pump is used to deliver the known volume... Simulated blood (from 0.1 mL to 20.0 mL in 0.1 mL increments) was added to a non-absorbent, diffuse-reflective standard background plate. After each drop stabilized, the system acquired an image, segmented the blood pool area, and calculated the area of ​​that region. Compared with average red saturation In the second stage, for the surgical dressing model, the known volume was... Simulated blood was evenly dripped onto a standard dry dressing in increments of 0.1 mL, ranging from 0.1 mL to 15.0 mL. The system acquired images, segmented the dressing area, and calculated its area. Compared with average red saturation After calibration, the system will collect all the data ( , Data points and their corresponding The truth values ​​are correlated and fitted to generate a pooled blood volume model. At the same time, ( , Data points and their Truth values ​​are correlated to generate a dressing absorption volume model. These two models are stored and used as the quantitative basis for the bleeding assessment unit.

[0047] While performing the first and second stage calibrations, the system reuses all collected image data for offline training of the target classification unit. The system's image feature extraction unit extracts the LBP texture histogram (texture features) and contour regularity parameters (contour features) from all pooled blood sample images and labels these feature vectors as pooled blood categories. Simultaneously, for all surgical dressing sample images, which cover the entire sequence from dry to fully saturated, the system extracts their LBP texture histograms and contour features and labels these feature vectors as surgical dressing categories. The system uses these two labeled feature vector datasets to train a support vector machine (SVM) classifier, which determines the decision boundaries for distinguishing between low-frequency fluid textures and high-frequency fabric textures, as well as irregular and regular contours. The trained SVM model is then embedded into the target classification unit, enabling it to distinguish between pooled blood and surgical dressings in subsequent actual operation.

[0048] Example 6: This example discloses a calibration procedure for determining the internal morphological parameters of a dynamic adaptive unit. This procedure is used to calibrate the area threshold used to identify the coagulation state. The initial state is as follows: under controlled illumination, i.e., a color temperature of 5500K and an illuminance of 1000 lux, two sets of image samples are acquired using a camera consistent with the system. Sample group A contains 200 images of simulated blood clots of different sizes, i.e., diameters ranging from 5mm to 20mm; sample group B contains 200 images of randomly dropped simulated blood drops with a diameter less than 2mm. The calibration process includes: running the first two steps of the coagulation state identification algorithm on all samples, i.e., obtaining and calculating the failure mask. Mask with dark red area And calculate the candidate condensation mask. Secondly, statistical sample group A and sample group B in The area distribution of connected components formed in the sample group is determined. Then, the distribution data is analyzed to determine a separation threshold that can eliminate more than 99% of the connected components of sample group B while retaining more than 95% of the connected components of sample group A. The area threshold is statistically determined to be 50 pixels. This procedure is also used to determine the size of the morphological opening kernel. By testing the elimination effect of different kernels, such as 1x1, 3x3, and 5x5, on sample group B, it is determined that the 3x3 kernel is sufficient to eliminate the remaining noisy connected components and has less than 2% morphological impact on sample group A. The calibration result is that the morphological opening kernel of the condensation state recognition algorithm is set to 3x3 and the connected component area threshold is set to 50 pixels.

[0049] It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, and that the present invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the present invention.

[0050] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention.

Claims

1. An intelligent intraoperative blood loss image assessment system, characterized in that, The system includes: The image feature extraction unit is used to extract texture and contour features from the acquired surgical area image; The target classification unit is used to identify target regions in surgical area images based on texture and contour features, and to distinguish target regions as collected blood or surgical dressings. The target tracking unit is used to initiate target tracking on the target area identified as surgical dressing by the target classification unit and predict the expected location of the target area in subsequent image frames. The evaluation logic control unit is used to determine the target area that falls within the expected location area as a surgical dressing; and the evaluation logic control unit is also used to bypass the target classification unit to re-determine the target area based on texture features when the target area is determined to be a surgical dressing because it falls within the expected location area. In addition, the illumination calibration unit also includes a baseline reset module; the baseline reset module is used to perform long-term analysis on the time series of image statistical features; the baseline reset module is also used to determine that the visual acquisition baseline has been systematically shifted when the value of the determined image statistical features continuously exceeds a preset offset threshold; the baseline reset module is also used to trigger the recalibration of the baseline parameters of the evaluation model in response to the systematic shift. The system also includes a dilution compensation unit; the dilution compensation unit includes a preset dilution compensation evaluation model; the dilution compensation unit is used to extract in parallel the second visual features in the surgical area image to characterize the presence of irrigation fluid, the second visual features being specular reflection features and transient deformation features; the dilution compensation unit is also used to generate a dynamic dilution mask covering the area affected by irrigation fluid based on the second visual features; the dilution compensation unit is also used to automatically call the dilution compensation evaluation model to calculate the bleeding volume for the area covered by the dynamic dilution mask.

2. The intelligent intraoperative blood loss image assessment system according to claim 1, characterized in that, The system also includes a bleeding volume assessment unit, which is used to calculate the blood volume of target areas classified as pooled blood and surgical dressings by the target classification unit based on the image features of the target area. The blood volume calculation includes: analyzing the red saturation and area of ​​the target area, and mapping the red saturation and area to a preset volume model to obtain the blood volume belonging to pooled blood and the blood volume belonging to surgical dressings, respectively.

3. The intelligent intraoperative blood loss image assessment system according to claim 2, characterized in that, The bleeding assessment unit includes an assessment model; the system also includes an illumination calibration unit, which is used to extract image statistical features representing illumination stability of surgical area images in parallel when assessing bleeding. The illumination calibration unit is also used to determine the credibility of the surgical area image for bleeding assessment based on image statistical features. The illumination calibration unit is also used to dynamically weight the bleeding assessment result sequence output by the assessment model within a time window according to the credibility.

4. The intelligent intraoperative blood loss image assessment system according to claim 3, characterized in that, The system also includes a dynamic adaptive unit; the dynamic adaptive unit includes a set of preset bleeding patterns. The dynamic adaptive unit is used to extract dynamic process features that characterize the rate of morphological expansion or color change of the target area based on a continuous sequence of surgical area images. The dynamic adaptive unit is also used to classify the current bleeding process into one of the preset bleeding patterns according to the dynamic process features. The dynamic adaptive unit is also used to adaptively adjust the internal parameters of the evaluation model based on the bleeding pattern.

5. The intelligent intraoperative blood loss image assessment system according to claim 4, characterized in that, The dynamic adaptive unit extracts dynamic process features of color change rate by calculating continuous image frames. and Color histograms of target areas categorized as blood collection or surgical dressings and The distance between the two Bachs This is achieved through the Bach distance. The calculation satisfies: ,in, and For continuous image frames and The color histogram of the target region. For the group index of the histogram, The total number of groups It is the natural logarithm function.

6. The intelligent intraoperative blood loss image assessment system according to claim 4, characterized in that, The dynamic adaptive unit is also used to obtain intermediate information generated by the dynamic process feature calculation when extracting dynamic process features. The intermediate information represents pixel areas where the calculation fails or the residual is too large. The dynamic adaptive unit is also used to identify the micro-state features of the target area classified as pooled blood based on the intermediate information. The micro-state feature is the coagulation state. The dynamic adaptive unit is also used to modify the evaluation logic of the evaluation model using the micro-state features. The modified evaluation logic will remove the areas that have been identified as coagulation state from the calculation of active bleeding volume.

7. The intelligent intraoperative blood loss image assessment system according to claim 3, characterized in that, Image statistical features include: histogram distribution features, which are used to quantify the pixel ratio of highlight areas and the pixel ratio of dark areas, and high-frequency information features, which are used to quantify the strong edge energy caused by specular reflection.

8. The intelligent intraoperative blood loss image assessment system according to claim 1, characterized in that, The evaluation logic control unit is also used to: count in real time the number of targets currently located in the surgical area image that are identified as surgical dressings by the target classification unit and are being tracked by the target tracking unit, and to compare the number of targets with the preset dressing baseline number.

Images