Model training and optimization method and system for blood loss estimation

Abstract

The present invention provides a model training and optimization method and system for blood loss estimation, the method comprising: automatically extracting first-class features from blood-stained gauze through a convolutional neural network; extracting artificially designed features as second-class features, the second-class features including the area of ​​the blood-stained region, the color moment in the RGB color space, and the color histogram in the HSV color space; converting the first-class features and the second-class features to a unified feature space through a feature conversion network, completing feature vector splicing, and obtaining fused features through a feature fusion network; and outputting the hemoglobin content and the amount of bleeding through a two-way regression network. The system comprises a feature extraction module, a feature conversion module, a feature fusion module, and a regression prediction module. The present invention improves the accuracy and reliability of blood loss estimation; combining deep learning with expert knowledge, it can provide strong support for medical decision-making and promote patient safety and health management.

Description

Technical Field

[0001] The present invention relates to the field of computer image processing and artificial intelligence technology, and in particular to a model training and optimization method and system for blood loss estimation. Background Art

[0002] During surgery or childbirth, accurate and timely assessment of blood loss is very important. On the one hand, underestimation of blood loss may lead to hemoglobin (Hb) deficiency caused by delayed blood transfusion, which in turn leads to circulatory instability and insufficient oxygen supply; on the other hand, overestimation of blood loss may lead to unnecessary waste of blood products and increase the morbidity and mortality of complications caused by blood transfusion. Accurate estimation of blood loss is crucial for timely initiation of treatment measures and avoiding unnecessary blood transfusion. However, there is currently a lack of simple, easy and accurate methods for blood loss assessment. How to assess blood loss in real time, dynamically and accurately has always been a research hotspot and clinical problem.

[0003] Prior art 1, application number: CN 202211682448.7 discloses a method and system for generating a perioperative blood transfusion plan based on a cascade model, including obtaining historical preoperative anemia correction data, establishing a preoperative anemia prediction model based on the historical preoperative anemia correction data; obtaining historical surgical data, establishing an intraoperative blood loss prediction model based on the historical surgical data, and obtaining intraoperative blood loss prediction information; obtaining historical intraoperative blood transfusion data, establishing an intraoperative blood transfusion prediction model based on the historical intraoperative blood transfusion data, and obtaining intraoperative blood transfusion prediction information; obtaining historical surgical records of intraoperative massive bleeding, establishing a massive bleeding prediction model based on the historical surgical records, and obtaining an intraoperative massive bleeding risk factor assessment report; although by establishing multiple models, multiple models are interconnected to generate perioperative preoperative, intraoperative and postoperative blood transfusion strategies, thereby providing a decision-making basis for the patient's blood transfusion plan. However, the blood loss prediction model requires a lot of sample data and is too dependent on the reliability of the blood loss prediction model, resulting in a long training time for the blood loss prediction model.

[0004] Prior art 2, application number: CN202410289556.0 discloses a MICU postpartum hemorrhage early warning system, including: a flow detection module, which is installed under the maternal mattress, for real-time detection of maternal bleeding and sending relevant data to the prediction model machine, so as to match with various physiological indexes of the maternal, and finally obtain the early warning level; a prediction model machine, which can establish a prediction model for postpartum hemorrhage through machine learning methods, and it measures and records the physiological indicators and bleeding volume of MICU critically ill patients in real time, extracts the clinical data such as the patient's high-risk factors for postpartum hemorrhage, hemoglobin and coagulation function, completes the establishment of the prediction model, and makes a judgment, and transmits the judgment result to the early warning module; an early warning module, which can identify MICU critically ill patients who are prone to postpartum hemorrhagic shock according to the prediction model machine, and issue a warning through sound or video, so that medical staff can take corresponding emergency measures in time. Although it can monitor and record postpartum physiological indicators and postpartum bleeding volume in real time, it is of great value for accurately assessing the patient's postpartum hemorrhage. However, the conventional machine learning method is used to establish a prediction model, in which the feature extraction method is relatively ordinary and overly dependent on historical data, which directly affects the training effect of the prediction model and causes the prediction results of the prediction model to be not very accurate.

[0005] Prior art three, application number: CN 202310738809.3 discloses a method and system for estimating and assessing the amount of postpartum hemorrhage based on machine learning. Through machine learning, images of bleeding volume of gauze, cloth, and the ground are obtained, and the bleeding-related conditions of the parturient are estimated through the bleeding volume images. At the same time, a risk assessment model is formed by obtaining and inputting the physical index information of the pregnant woman, such as blood pressure, pulse, respiration, body temperature, blood oxygen saturation, hemoglobin, etc., to clarify and predict the risk level of the pregnant woman. According to the principle of three-level early warning measures for postpartum hemorrhage, targeted prevention or treatment measures are given. Although the accuracy of the estimation of the amount of bleeding is further improved, the detection rate of postpartum hemorrhage is improved, the underestimation rate of postpartum hemorrhage and the incidence rate of severe postpartum hemorrhage are effectively reduced, and the amount of postpartum hemorrhage can be estimated more quickly and accurately, which is of great significance for reducing the incidence and mortality rate of postpartum hemorrhage in parturients. However, a large amount of sample data is used for learning, and there is a lack of support for sample data involved manually, resulting in the reliability and robustness of the sample data to be further improved.

[0006] At present, the prior art 1, prior art 2 and prior art 3 have a single method for obtaining technical features and lack of integration with artificially designed technical features, resulting in the problem that the accuracy and consistency of the existing blood loss estimation methods are poor. Therefore, the present invention provides a model training and optimization method and system for blood loss estimation, which makes full use of artificial features and automatic learning features, and extracts more complete color information related to hemoglobin in blood-stained gauze through deep learning methods, and more accurately estimates the blood loss and hemoglobin content, thereby overcoming the shortcomings of the existing blood loss estimation methods, which are poor in accuracy and consistency. Summary of the invention

[0007] In order to solve the above technical problems, the present invention provides a model training and optimization method and system for blood loss estimation, comprising the following steps:

[0008] The first type of features are automatically extracted through a convolutional neural network; the artificially designed features are extracted as the second type of features, and the second type of features include the area of ​​the blood-stained region, the color moment in the RGB color space, and the color histogram in the HSV color space;

[0009] The first and second features are converted to a unified feature space through a feature conversion network, the feature vectors are concatenated, and the converted features are passed through a feature fusion network to obtain fused features;

[0010] After obtaining the fused features, the fused features are output through two fully connected networks to output the hemoglobin content and bleeding volume respectively.

[0011] The process of obtaining the first type of features includes the following steps:

[0012] Obtaining a blood-soaked gauze sample image, inputting the blood-soaked gauze sample image into a convolutional layer of a convolutional neural network, wherein a convolution kernel in the convolutional layer slides on the input blood-soaked gauze sample image, multiplies the image point by point with a local area of ​​the blood-soaked gauze sample image, and adds the results to generate a new feature map;

[0013] Optionally, a residual unit is added to the convolutional neural network by introducing a short-circuit mechanism to learn the difference between the blood-soaked gauze sample image at the input and output of the convolutional neural network, thereby retaining the input information of the blood-soaked gauze sample image;

[0014] The output convolutional neural network automatically learns to extract the first-class features of the blood-soaked gauze sample.

[0015] Optionally, the convolutional neural network adopts a ResNet-101 structure as a convolutional neural network, consisting of 1 conv1 layer, 3 conv2_x blocks, 4 conv3_x blocks, 23 conv4_x blocks, 3 conv5_x blocks and one fc layer, with a total of 1+3*(3+4+23+3)+1=101 layers.

[0016] The process of obtaining the second type of features includes the following steps:

[0017] Obtain a sample image of blood-soaked gauze, identify the red area on the gauze, and divide the number of pixels in the red area by the total number of pixels in the image as the area feature;

[0018] In the RGB space, find the blood-stained area in the blood-soaked gauze sample image, and divide the number of pixels in the blood-stained area by the total number of pixels in the image as the area feature. Among them, the pixels in the blood-stained area meet the following conditions:

[0019] 1<R / G<U limit

[0020] Among them U limit is the upper threshold limit of the ratio. The reason for adopting this condition is that in the RGB color space, the R, G, and B components of the blood pixel satisfy: R>G>B, U limit The value of is set to [2, 4, 8, 16, 32, 64, 128, 256, +∞], and 9 red intervals are obtained. The pixels in all red intervals that satisfy B>G are classified into the 10th category, and a total of 10 area features can be obtained;

[0021] In RGB space, take the image mean E in the three color channels i , standard deviation σ i and skewness i As color moment features, color moment features reflect the spatial distribution characteristics of colors;

[0022] The blood-soaked gauze sample image is converted from RGB space to HSV space, and H, S, and V are quantized into multiple statistical bins respectively. For each pixel in the blood-soaked gauze sample image, its corresponding quantization level is found and counted in the corresponding statistical bin to obtain a multidimensional vector representing the number or frequency of pixels in each quantization bin.

[0023] Optionally, the H, S, and V channels are quantized at 12, 8, and 8 levels respectively, and a multidimensional histogram is generated using the three quantized channels, thereby generating a histogram containing 768 statistical bins.

[0024] Optionally, the jth pixel value of the i-th color channel is p i,j , the three color moments are expressed as follows:

[0025]

[0026] Where N is the number of pixels, and 9 color moment features are obtained from 3 color channels in the color space.

[0027] The process of obtaining fused features through the feature fusion network includes the following steps:

[0028] Obtain four features: 768 color histogram features, 10 area features, 9 color moment features, and 1024 convolution features;

[0029] A conversion network consisting of four single-layer or multi-layer fully connected networks is used to transform the four features into a unified feature space;

[0030] Through a two-layer fully connected network, the transformed features in the feature space are fused to obtain a 128-dimensional feature vector.

[0031] The present invention provides a model training and optimization system for blood loss estimation, comprising:

[0032] The feature extraction module is responsible for obtaining the pre-made blood-soaked gauze sample and automatically learning and extracting the first type of features of the blood-soaked gauze sample through the convolutional neural network; extracting the artificially designed features as the second type of features, which include the blood-stained area, the color histogram features in the HSV color space, and the color moment in the RGB color space;

[0033] The feature conversion module is responsible for converting the first and second features into a unified feature space through the feature conversion network and completing the splicing of feature vectors;

[0034] Feature fusion module, responsible for obtaining fusion features through feature fusion network;

[0035] The regression prediction module takes the fused features as input and outputs the hemoglobin content and bleeding volume respectively through two fully connected networks.

[0036] The present invention provides basic training data for the model by collecting pre-made blood-soaked gauze samples. The samples are the key to the model's identification and evaluation of blood loss. The representative training samples can ensure that the model obtains a variety of features and data distribution information during the learning process, thereby improving the generalization ability of the model. The first type of features are deep features automatically learned and extracted by convolutional neural networks (CNNs). CNNs can automatically extract rich high-dimensional features from image data, capture the complex structure, color information and spatial relationship of the image, and improve the accuracy of blood loss estimation. Using CNN's deep feature extraction enables the algorithm to learn rich feature representations from the original data without the need for artificial feature design. The second type of features includes blood-stained area features, color histogram features extracted by HSV color space, and color moment features in RGB color space. The second type of features contains expert domain knowledge, based on the features of blood-stained gauze color information (such as the color characteristics of red blood) and the physical properties of the image, and supplements the deficiencies of the machine learning model with artificially designed features, which helps to improve the accuracy of the model in the process of blood loss assessment, especially the adaptability under different environments and lighting conditions. The first and second features are mapped to a unified feature space through a feature conversion network, and then the feature vectors are spliced; feature fusion can combine the features automatically learned by the deep learning model with the manually designed domain knowledge to form a more comprehensive feature description and improve the expressiveness of the features; the feature fusion network further processes the spliced ​​features to extract more abundant and discriminative fusion features; fusion features can increase the expressiveness of the model, so that the blood loss estimation not only depends on the recognition result of a single feature, but also integrates the information of multiple features, thereby improving the accuracy and reliability of the prediction. Through two fully connected networks, the hemoglobin content and bleeding volume are output respectively, and the modeling of the two target variables is clearly distinguished to ensure that each output part can focus on its specific prediction task; the prediction of hemoglobin content and bleeding volume can provide more detailed clinical data support, and provide doctors with accurate data basis when evaluating the condition and formulating treatment plans; the use of fusion features can better capture the complexity of the blood loss estimation process, such as the relationship between the color change of blood and the amount of bleeding; accurate estimation of blood loss is crucial for clinical emergency and patient monitoring, which can help doctors take timely intervention measures to reduce the risk of complications and ensure the safety of patients.

[0037] Other features and advantages of the present invention will be described in the following description, and partly become apparent from the description, or understood by practicing the present invention. The purpose and other advantages of the present invention can be realized and obtained by the structures particularly pointed out in the written description and the accompanying drawings.

[0038] The technical solution of the present invention is further described in detail below through the accompanying drawings and embodiments. BRIEF DESCRIPTION OF THE DRAWINGS

[0039] The accompanying drawings are used to provide a further understanding of the present invention and constitute a part of the specification. Together with the embodiments of the present invention, they are used to explain the present invention and do not constitute a limitation of the present invention. In the accompanying drawings:

[0040] Figure 1 This is a flow chart of a model training and optimization method for estimating blood loss in Example 1 of the present invention;

[0041] Figure 2 This is a diagram of the process of obtaining the first type of features in Example 2 of the present invention;

[0042] Figure 3 This is a diagram of the process of obtaining the second feature in Embodiment 3 of the present invention;

[0043] Figure 4 This is a process diagram of obtaining fused features through a feature fusion network in Example 4 of the present invention;

[0044] Figure 5 This is a schematic diagram of the model structure in Example 4 of the present invention;

[0045] Figure 6 This is a process diagram of outputting hemoglobin content and bleeding volume in Example 5 of the present invention;

[0046] Figure 7 This is a block diagram of a model training and optimization system for estimating blood loss in Example 6 of the present invention. DETAILED DESCRIPTION

[0047] The preferred embodiments of the present invention are described below in conjunction with the accompanying drawings. It should be understood that the preferred embodiments described herein are only used to illustrate and explain the present invention, and are not used to limit the present invention.

[0048] The terms used in the embodiments of the present application are only for the purpose of describing specific embodiments, and are not intended to limit the embodiments of the present application. The singular forms of "a", "said" and "the" used in the embodiments of the present application are also intended to include plural forms, unless the context clearly indicates other meanings. It should also be understood that the term "and / or" used in this article refers to and includes any or all possible combinations of one or more associated listed items.

[0049] When the following description relates to the accompanying drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with the present application. On the contrary, they are only examples of devices and methods consistent with some aspects of the present application. In the description of the present application, it should be understood that the terms "first", "second", "third", etc. are only used to distinguish similar objects, and do not have to be used to describe a specific order or sequence, nor can they be understood as indicating or implying relative importance. For those of ordinary skill in the art, the specific meanings of the above terms in the present application can be understood according to the specific circumstances.

[0050] Example 1: Figure 1 As shown, an embodiment of the present invention provides a model training and optimization method for blood loss estimation, comprising the following steps:

[0051] S100: obtaining a pre-made blood-soaked gauze sample image, and automatically learning and extracting the first type of features of the blood-soaked gauze sample through a convolutional neural network; extracting artificially designed features as the second type of features, the second type of features including the blood-stained area, the color histogram features in the HSV color space, and the color moment;

[0052] S200: converting the first type of features and the second type of features into a unified feature space through a feature conversion network, completing the concatenation of feature vectors, and obtaining fused features through a feature fusion network;

[0053] S300: Obtain fusion features and output the hemoglobin content and bleeding volume respectively through two fully connected networks.

[0054] The working principle and beneficial effects of the above technical solution are as follows: the present embodiment first obtains a pre-made blood-soaked gauze sample image, and automatically learns and extracts the first type of features of the blood-soaked gauze sample through a convolutional neural network; at the same time, the blood-stained area, the color histogram in the HSV color space, and the color moment in the RGB color space are extracted as the second type of features, and the second type of features are artificially designed features containing expert domain knowledge; secondly, the first type of features and the second type of features are respectively converted to a unified feature space through four fully connected networks of the feature conversion network to complete the splicing of feature vectors. The fused features are obtained through the feature fusion network; finally, the fused features are obtained, and the hemoglobin content and the amount of bleeding are output respectively through two fully connected networks. Step S100 of the above scheme provides basic training data for the model by collecting pre-made blood-soaked gauze sample images. The sample is the key to the model's identification and evaluation of blood loss; representative training samples can ensure that the model obtains a variety of features and data distribution information during the learning process, thereby improving the generalization ability of the model. The first type of feature automatically learns to extract deep features through convolutional neural network (CNN). CNN can automatically extract rich high-dimensional features from the blood-soaked gauze sample image data, capture the complex structure, color information and spatial relationship of the blood-soaked gauze sample image, and improve the accuracy of blood loss estimation. The deep feature extraction of CNN enables the algorithm to learn rich feature representations from the original data without artificial feature design. The second type of feature extracts the blood-stained area feature, the color histogram feature of the HSV color space, and the color moment feature of the RGB space. The second feature contains expert domain knowledge, based on the features of the blood-soaked gauze color information (such as the color characteristics of red blood) and the physical properties of the blood-soaked gauze sample image. The artificially designed features supplement the deficiencies of the machine learning model, which helps to improve the accuracy of the model in the process of blood loss assessment, especially the adaptability under different environments and lighting conditions. Step S200 maps the first and second features to a unified feature space through a feature conversion network, and then concatenates the feature vectors; feature fusion can combine the features automatically learned by the deep learning model with the manually designed domain knowledge to form a more comprehensive feature description and improve the expressiveness of the features; the feature fusion network further processes the concatenated features to extract richer and more discriminative fusion features; fusion features can increase the expressiveness of the model, so that the blood loss estimation not only depends on the recognition result of a single feature, but integrates the information of multiple features, thereby improving the accuracy and reliability of the prediction.Step S300 outputs the hemoglobin content and the amount of bleeding respectively through two fully connected networks, clearly distinguishing the modeling of the two target variables and ensuring that each output part can focus on its specific prediction task; separately predicting the hemoglobin content and the amount of bleeding can provide more detailed clinical data support, providing doctors with accurate data basis when evaluating the condition and formulating treatment plans; using fusion features can better capture the complexity of the blood loss estimation process, such as the relationship between the color change of blood and the hemoglobin content; accurate estimation of blood loss is crucial for clinical emergency and patient monitoring, and can help doctors take timely intervention measures to reduce the risk of complications and ensure patient safety.

[0055] This embodiment applies machine vision and artificial intelligence technology to the estimation of bleeding volume to obtain a new method for estimating bleeding volume; it can overcome the shortcomings of poor accuracy, inconvenient operation, and delayed results in the existing bleeding volume estimation methods; it meets the requirements of simple, timely and accurate bleeding volume estimation. Before the emergence of convolutional neural networks, the color features of blood-soaked gauze sample images depended on the knowledge of domain experts, and different color features were manually designed and extracted, including hue, saturation, brightness, color moments, chromaticity moments, color correlograms, etc. After the emergence of convolutional neural networks, color features can be automatically learned from data. Compared with the manual method, the learning method can obtain more comprehensive and effective features. With the help of effective features and a large number of labeled samples, the amount of blood loss can be quickly and effectively calculated automatically through deep learning technology. Although deep convolutional neural networks can automatically learn good features from data, expert knowledge and manually designed features are still of great value. Unbalanced distribution and noise of training data will lead to deviations in automatically learned features; insufficient training data will lead to poor performance of learned feature representation; insufficient number of convolution kernels and unreasonable design will lead to incomplete learned feature representation. Manually designed features reflect the professional knowledge of experts in the field of blood-soaked gauze sample images. They are reliable, effective, and robust, and can make up for the shortcomings of automatically learned features. In order to combine the advantages of two different features, an effective method is needed to effectively fuse learned features and manually designed features to obtain a more complete feature representation with complementarity. Feature fusion is the process of processing different types of feature vectors extracted from the same pattern to obtain optimized feature vectors. It requires more discriminative power than any original single input feature vector, and utilizes the complementarity between features to fuse the advantages between features, thereby improving the performance of the model. In the blood-soaked gauze sample image data, there are two most basic methods of feature fusion: point-wise addition and vector concatenation. The two types of features in the present invention are automatically learned features and manually designed features. The feature spaces and feature dimensions of these two types of features are very different, and are not suitable for direct use of the add or concatenate method. The importance of each feature is automatically obtained by learning, and then the features that are more useful for regression performance are improved according to this importance, while suppressing the features that are not very useful for other regression tasks. For this purpose, a feature conversion network and a feature fusion network are proposed. The feature conversion network is used to convert different types of features into a unified feature space through a fully connected network, and then perform feature vector splicing.The feature fusion network embeds the concatenated vector into the final fusion feature space through a fully connected network.

[0056] In summary, the entire model training and optimization method of this embodiment cooperates through multiple steps, and the design of each step is centered on improving the accuracy and reliability of blood loss estimation; through efficient feature extraction and fusion, combined with deep learning and expert knowledge, it can provide strong support for medical decision-making and promote patient safety and health management.

[0057] Example 2: Figure 2 As shown, based on Example 1, the process of obtaining the first type of features provided by the embodiment of the present invention includes the following steps:

[0058] S101: obtaining a blood-soaked gauze sample image, inputting the blood-soaked gauze sample image into a convolutional layer of a convolutional neural network, sliding a convolution kernel in the convolutional layer on the input blood-soaked gauze sample image, performing point-by-point multiplication with a local area of ​​the blood-soaked gauze sample image, and adding the results to generate a new feature map;

[0059] S102: adding a residual unit to the convolutional neural network by introducing a short-circuit mechanism, learning the difference between the blood-soaked gauze sample image in the convolutional neural network input and output, and retaining the input information of the blood-soaked gauze sample image;

[0060] S103: Output the convolutional neural network to automatically learn and extract the first type of features of the blood-soaked gauze sample.

[0061] The working principle and beneficial effects of the above technical solution are as follows: the present embodiment first obtains a sample image of blood-soaked gauze, inputs the sample image of blood-soaked gauze into the convolution layer of the convolution neural network, and the convolution kernel in the convolution layer slides on the input sample image of blood-soaked gauze, multiplies point by point with the local area of ​​the sample image of blood-soaked gauze, and adds the results to generate a new feature map; secondly, the convolution neural network introduces a short-circuit mechanism to add a residual unit, learns the difference between the input and output of the sample image of blood-soaked gauze in the convolution neural network, and retains the input information of the sample image of blood-soaked gauze; finally, the output convolution neural network automatically learns to extract the first feature of the sample of blood-soaked gauze. Step S101 of the above scheme obtains a sample image of blood-soaked gauze and generates a feature map. Through the sliding of the convolution layer and the convolution kernel, the network can extract local features, including edges, shapes and colors, etc., from the input sample image of blood-soaked gauze; each convolution kernel performs a convolution operation on the input sample image of blood-soaked gauze to generate a feature map, which shows the response of the sample image of blood-soaked gauze to the specific features (for example, a certain shape or color) that the convolution kernel focuses on. Significance achieved: Through local feature extraction, a large amount of information is compressed into feature maps, providing more concentrated and useful information, significantly reducing the computational complexity; multiple convolution kernels can learn different features, improving the model's ability to understand and classify complex blood-soaked gauze sample images. Step S102 introduces residual units and short-circuit mechanisms. Through short-circuit connections, the network can directly learn the difference between input data and output, simplifying the learning task of the model; the design of residual units enables the output of multiple layers to retain input information, reducing information loss in deep networks. Significance achieved: Deep neural networks often encounter the problem of gradient disappearance or explosion. The use of short-circuit connections can provide a better gradient flow path and improve the training effect; the introduction of residual units allows the network to accommodate more layers without performance degradation, allowing the model to capture more complex features. Step S103 outputs the first type of features. The convolutional neural network automatically learns the first type of features of the blood-soaked gauze sample through a combination of multiple convolutional layers and residual units; the output first type of features provides a solid foundation for subsequent layer processing and further learning, and can perform feature conversion and fusion at a higher level. Significance achieved: High-quality first-class features are the basis for good network performance, which can help the model understand the pattern of input data more accurately and effectively process blood-soaked gauze samples; the extracted first-class features lay the foundation for subsequent regression analysis tasks, so that the model can maintain accuracy and improve robustness.

[0062] In summary, the role of each step in the feature acquisition process of this embodiment complements each other, forming an effective feature extraction framework. By extracting local features through convolutional layers and retaining information and alleviating gradient challenges through residual units, efficient feature learning is achieved, and the first type of features finally output can significantly improve the performance and efficiency of subsequent tasks; the idea of ​​hierarchical feature extraction is one of the important advantages of deep learning, especially convolutional neural networks.

[0063] The convolutional neural network of this embodiment can automatically extract local features in the blood-soaked gauze sample image, and combine low-level features into high-level feature representations by layer-by-layer transmission. The convolution kernel in the convolution layer slides on the input blood-soaked gauze sample image, multiplies point by point with the local area of ​​the blood-soaked gauze sample image, and adds the results to generate a new feature map. This operation can extract local features in the blood-soaked gauze sample image, such as edges, shapes, and colors. By adjusting the parameters of the convolution kernel, under the guidance of the loss function, CNN can learn different feature representations. The ResNet-101 structure is used as a convolutional neural network, which consists of 1 conv1 layer, 3 conv2_x blocks, 4 conv3_x blocks, 23 conv4_x blocks, 3 conv5_x blocks and an fc layer, with a total of 1+3*(3+4+23+3)+1=101 layers. ResNet is an excellent and widely used CNN feature extraction network. ResNet adds residual units to CNN through a short-circuit mechanism, and solves the network degradation and gradient vanishing problems by learning residuals, so that ResNet can train deeper networks from hundreds to thousands of layers and achieve higher accuracy. Another important reason for choosing RetNet101 is that the amount of blood loss is mainly related to the color of the blood-soaked gauze sample image, and has little to do with texture and shape information, so the color features of the blood-soaked gauze sample image are focused on. The color of the blood-soaked gauze sample image comes from the combination of different channel components of the blood-soaked gauze sample image pixels, so in the convolutional network, more attention is paid to the information interaction between channels. The 1x1 convolution of Bottleneck Architectures in the building block of Resnet-101 plays such a role.

[0064] For the i-th training sample, assume is the blood volume, is the hemoglobin content, and the predicted output blood loss and hemoglobin concentration are v i and m i , the loss function is taken as:

[0065]

[0066] Where N is the number of samples, α and β are the weight coefficients of the loss function.

[0067] Example 3: Figure 3 As shown, based on Example 1, the process of obtaining the second type of features provided by the embodiment of the present invention includes the following steps:

[0068] S104: Obtain a blood-soaked gauze sample image, convert the blood-soaked gauze sample image from RGB to an HSV space model, quantize the H, S, and V channels at 12, 8, and 8 levels, respectively, and use the three quantized channels to generate 768 statistical bins. For each pixel in the blood-soaked gauze sample image, find its corresponding quantization level, count it in the corresponding statistical bin, and obtain a 768-dimensional vector, representing the number or frequency of pixels in each quantization bin;

[0069] S105: Find the blood-stained area in the blood-soaked gauze sample image in the RGB space, and divide the number of pixels in the blood-stained area by the total number of pixels in the image as the area feature;

[0070] Among them, the pixels in the blood-stained area meet the following conditions:

[0071] 1<R / G<U limit

[0072] Among them U limit is the upper threshold limit of the ratio. The reason for adopting this condition is that in the RGB color space, the R, G, and B components of the blood pixel satisfy: R>G>B;

[0073] When the gauze is used to wipe the blood, the blood will spread to the surrounding area on the gauze, and the concentration of the blood will gradually fade from the center to the surrounding area. It is not enough to use a single red area to represent this feature, so the red area is divided into 10 levels. limit The value of is set to [2, 4, 8, 16, 32, 64, 128, 256, +∞], and 9 red intervals are obtained. Due to the influence of ambient lighting changes and noise, there may be a situation where the color component B>G in the red area. All pixels in the red interval that satisfy B>G are classified into the 10th category, and a total of 10 area features can be obtained;

[0074] S106: extracting a color histogram of the blood-soaked gauze sample image, wherein the color moment includes a mean, a standard deviation, and a skewness; and obtaining 9 color moment features in three color channels;

[0075] Among them, assuming that the j-th pixel value of the i-th color channel is p i,j , the three color moments are expressed as follows:

[0076]

[0077] Where N is the number of pixels, and 9 color moment features can be obtained from 3 color channels in the color space.

[0078] The working principle and beneficial effects of the above technical solution are as follows: First, the blood-soaked gauze sample image is obtained, and the blood-soaked gauze sample image is converted from RGB to HSV space model, and the H, S, and V channels are quantized at 12, 8, and 8 levels respectively, and 768 statistical bins are generated using the three quantized channels; secondly, for each pixel in the blood-soaked gauze sample image, its corresponding quantization level is found, and counted in the corresponding statistical bin to obtain a 768-dimensional vector, representing the number of pixels or frequency in each quantization bin; in the RGB space, the blood-stained area in the blood-soaked gauze sample image is found, and the number of pixels in the blood-stained area is divided by the total number of pixels in the image as the area feature. Finally, the color histogram of the blood-soaked gauze sample image is extracted, and the mean, standard deviation, and skewness are used as color moments to obtain 9 color moment features in the three color channels. Step S104 of the above scheme acquires the image and performs preliminary processing. By capturing the image of the blood-soaked gauze sample as the basic data for subsequent analysis, the authenticity and validity of the data source are ensured; by converting the RGB image to the HSV color space, the color information can be processed more intuitively. Compared with RGB, the HSV space is easier to capture the characteristics of color, especially when processing images under different lighting conditions; the three channels of H, S, and V are quantized to 12, 8, and 8 levels respectively, and the color values ​​are discretized into reasonable intervals, providing a basis for histogram generation; quantization reduces the computational complexity and organizes the color information into a more processable form; a histogram containing 768 statistical bins is created, so that the frequency of occurrence of each color can be quantified in the subsequent process, and the relationship between different color dimensions is maintained. Significance: It can effectively extract color-related features and prepare for blood testing and analysis. The histogram is counted and the frequency vector is obtained. By searching each pixel value, it is mapped to the corresponding quantization level to ensure that each pixel can correctly correspond to the statistical bin of the histogram; the number of pixels in each quantization bin is counted, so as to update the frequency or number of pixels in each bin, and a 768-dimensional vector is generated, which clearly represents the distribution of each color quantization level in the image. Significance: The color information is abstracted from the image to form a feature vector that meets the needs of subsequent analysis, which is helpful for the comparison, classification and identification of different image samples, and improves the efficiency of subsequent machine learning models. Step S105 extracts the color area. By calculating the number of red pixels in the image, the blood staining and coverage in the image can be effectively captured, and the feature information is further enriched. Significance: There is a positive correlation between the area of ​​the blood-stained area and the amount of blood loss, which directly reflects the amount of blood loss. The blood-stained area of ​​the same area has different color shades, and the blood loss and hemoglobin content reflected are also different. Dividing the area of ​​the blood-stained area into multiple levels according to the color shade can well capture this difference.Step S106 extracts color moment features, which, combined with the histogram color features obtained previously, can reflect the pixel distribution in subsequent analysis; by calculating these three color moment features, the center position (mean) of the color distribution, the discrete degree (standard deviation) of the color value, and the symmetry (skewness) of the color distribution can be characterized. The multi-dimensional feature extraction method enhances the deep understanding of color information and makes the features more comprehensive; Significance: It integrates the rich color information in the image, and through quantification and statistics, a comprehensive description of the color distribution can be formed, which can improve the accuracy and efficiency of the analysis.

[0079] The red color of blood is caused by hemoglobin in red blood cells, so we are only interested in the red area in the gauze. The area of ​​the red area in the gauze and the brightness and purity of the red color determine the blood content and hemoglobin content. The artificially designed features include: the area of ​​the red area, the color histogram of the red area, and the color moment of the red area. The area of ​​the red area can be expressed as the number of pixels in the area divided by the total number of pixels in the blood-soaked gauze sample image.

[0080] The color histogram describes the proportion of different colors in the entire blood-soaked gauze sample image, and does not care about the spatial position of each color. It has the advantages of translation, rotation, and scale invariance, and can better represent the color information in the blood-soaked gauze sample image. The disadvantage of the color histogram of the blood-soaked gauze sample image is that the data volume is too large. For the 24-bit RGB blood-soaked gauze sample image, the number of color histogram feature statistical bins is as high as 256×256×256=16777216. Even if the color value of the blood-soaked gauze sample image is quantized to the range of 16×16×16, the number of color histogram feature statistical bins is 4096, which is still too large for the classifier. For this reason, the blood-soaked gauze sample image is converted from RGB space to HSV space. The three components of HSV space represent color (Hue), saturation (Saturation) and value (Value). H, S, and V are quantized to the range of 12*8*8, respectively, with a total of 768 statistical bins. Therefore, for a blood-soaked gauze sample image, a 768-dimensional global histogram feature can be obtained. The histogram only describes the color information of the blood-soaked gauze sample image, and loses the spatial distribution information of the color; to address this deficiency, the color moment is used to describe the distribution characteristics of the color in the color space. The basic idea of ​​the color moment is to assume that the color distribution of the blood-soaked gauze sample image can be interpreted as a certain probability distribution, and the probability distribution can be uniquely characterized by its moment, such as the normal distribution can be completely represented by its mean and standard deviation. Correspondingly, the color distribution can be characterized by the color moment. The first to third order central moments are used to represent the color moments of the blood-soaked gauze sample image, which are the mean, standard deviation and skewness, respectively. The skewness can be understood as a measure of the degree of asymmetry of the distribution.

[0081] In summary, the entire second-class feature acquisition process of this embodiment realizes the structured extraction of color information by systematically processing image data. Each step complements each other, so that the feature vector finally obtained can not only reflect the basic color information of the image, but also describe the characteristics of color distribution, which has important practical application value for the subsequent blood detection and classification tasks. In this way, the model can more accurately identify and analyze the color and morphological characteristics of blood in gauze, thereby achieving higher recognition rate and accuracy.

[0082] Example 4: Figure 4 As shown, based on Example 1, the process of obtaining fusion features through a feature fusion network provided in the embodiment of the present invention includes the following steps:

[0083] S201: Obtain four features, namely, 768 color histogram features, 10 area features, 9 color moment features, and 1024 convolution features;

[0084] S202: using a conversion network consisting of four single-layer or multi-layer fully connected networks to convert the four features into a unified feature space;

[0085] S203: Through a two-layer fully connected network, the four features in the feature space are fused to obtain a 128-dimensional feature vector.

[0086] The working principle and beneficial effects of the above technical solution are as follows: the present embodiment first obtains four features, namely, 768 color histogram features, 10 area features, 9 color moment features and 1024 convolution features; secondly, a conversion network consisting of four single-layer or multi-layer fully connected networks is used to convert the four features into a unified feature space; finally, a 128-dimensional feature vector is obtained by fusion of the four features in the feature space through a two-layer fully connected network. Step S201 of the above scheme obtains four features, and extracts four different types of features from an image or sample, namely, 768 color histogram features, 10 area features, 9 color moment features and 1024 convolution features. This diversity can help capture different attributes of a sample, such as color distribution, area information and deep features; each feature represents information about a sample in a specific aspect, and through these features, a multi-angle understanding of the sample can be obtained. Significance: By obtaining different types of features, the comprehensiveness of the model analysis can be enhanced, thereby improving the accuracy and robustness of subsequent tasks; the diversified features will provide sufficient information support for the subsequent feature conversion and fusion process. Step S202: Feature conversion, which converts different types of features (color histogram features, area features, color moment features, and convolutional features) into a unified feature space through multiple single-layer / multi-layer fully connected networks, is implemented by four different conversion networks, so that various types of features can be processed in the same way; the conversion network can convert high-dimensional features into low-dimensional feature representations through appropriate connection matrices, and can also map low-dimensional features to high-dimensional space. Significance: The unified feature space lays the foundation for feature fusion, allowing different features to be weighted and integrated at the same scale, enhancing the representation ability of feature information; converting features into a more unified form is helpful for subsequent computing tasks. Step S203: Feature fusion, which fuses the converted 512-dimensional features through two layers of fully connected networks, and finally obtains a 128-dimensional feature vector. This process integrates the information encoded by different types of features; due to the use of a fully connected network, the model can learn the complex relationship between features, so as to better extract comprehensive features. Significance: The fused feature vector can more effectively represent the overall characteristics of the sample, which helps to improve the performance of the model in subsequent regression tasks; the 128-dimensional fused features are the basis for the final decision of the model, which can enhance the ability to understand complex data and thus improve the prediction accuracy of targets (such as hemoglobin content and bleeding volume).

[0087] In this embodiment, after the feature extraction stage, four types of features are obtained, 768 color histogram features, 10 area features, 9 color moment features and 1024 convolution features. The four features are effectively fused. First, four conversion networks are used to convert the four features into a unified feature space. Assume that the four types of features are:

[0088] (X 1) 1x10

[0089] (X 2 ) 1x9

[0090] (X 3 ) 1x768

[0091] (X 4 ) 1x1024

[0092] The conversion network is a single-layer fully connected network, and the network connection matrix is:

[0093] (W 1 ) 10x64

[0094] (W 2 ) 9x64

[0095] (W 3 ) 768x128

[0096] (W 4 ) 1024x256

[0097] After feature conversion, the four features are concatenated to obtain a 512-dimensional feature vector:

[0098] F=[X 1 W 1 ,X 2 W 2 ,X 3 W 3 ,X 4 W 4 ] 1x512

[0099] It should be pointed out that the feature conversion network is not limited to a single-layer fully connected network. If the computational complexity is not considered, a multi-layer network can also be used.

[0100] The function of feature conversion is to embed four different features into a unified feature space, and then splice them into a unified feature. The spliced ​​features are passed through the feature fusion network to obtain the final fusion feature. The feature fusion network is a two-layer fully connected network. The connection matrix of the first layer network is (W f ) 512x256 , the second layer network connection matrix is ​​(W s ) 256x128 , after feature fusion, a 128-dimensional feature vector is obtained.

[0101] (F) 1x512 (Wf ) 512x256 (W s ) 256x128

[0102] The amount of blood loss and hemoglobin content are closely related. After obtaining the hemoglobin content, the amount of bleeding can be calculated by the hemoglobin concentration in human blood. Affected by factors such as flushing fluid and disinfectant, the hemoglobin concentration on the gauze is not necessarily the same as that in the human body. The hemoglobin concentration in the human body is not fixed and will decrease with blood loss. By using the learning ability of machine learning combined with pre-made learning samples, the hemoglobin content and bleeding volume can be calculated at the same time, which can achieve higher accuracy of bleeding volume.

[0103] The network structure of the entire model can be divided into four parts, namely feature extraction module, feature conversion module, feature fusion module, blood loss prediction network Blood head and hemoglobin prediction network HB head. The model structure is as follows Figure 5 As shown in the figure, the model uses the ResNet backbone neural network to automatically extract features from the blood-soaked gauze image. The automatic features and manual features are converted through the feature conversion network to obtain the converted features. The converted features are then converted through the feature fusion network to obtain the fused features. The fused features are then output through two fully connected networks to output the hemoglobin content and the amount of bleeding respectively. Since the hemoglobin content and the amount of bleeding are positively correlated, the Blood head and HB head share the feature extraction module, feature conversion module and feature fusion module, which can achieve better results than outputting one path separately.

[0104] The present invention utilizes conversion networks and fusion networks to obtain more complete fusion features with stronger representation capabilities through learning. The feature conversion network converts the color histogram features, bloodstained area features, color moment features and convolution features into a unified feature space through four fully connected networks, and completes the splicing of feature vectors in the unified feature space. In order to enhance the importance of features that are more useful for regression performance and suppress features that are less useful for other regression tasks, the feature vectors are spliced ​​and the final fused feature vector is obtained through a fully connected network. The connection parameters of the feature conversion network and the feature fusion network are automatically adjusted during the learning process, and more effective fusion features for regression tasks can be obtained through learning.

[0105] In summary, this embodiment jointly constructs a complete feature conversion and feature fusion network; by acquiring diverse features, mapping them to a unified feature space, and fusing features through a fully connected network, the model can effectively integrate and process complex information. It can significantly improve the accuracy and value of medical sample analysis and provide a more accurate basis for clinical decision-making. The feature fusion process not only improves the performance of the model, but also opens up new directions for deeper medical research and applications.

[0106] Example 5: Figure 6 As shown, based on Example 1, the process of outputting hemoglobin content and bleeding volume provided in the embodiment of the present invention comprises the following steps:

[0107] S301: constructing a blood loss prediction network and a hemoglobin prediction network including an input layer, a first hidden layer, a second hidden layer and an output layer;

[0108] S302: The fused feature, i.e., the 128-dimensional feature vector, is passed as input to the two-way fully connected network, the first hidden layer maps the 128-dimensional vector to 64-dimensional, the second hidden layer maps the 64-dimensional vector to 128-dimensional, and the output layer maps the 128-dimensional vector to a scalar for predicting blood loss;

[0109] S303: The input also comes from the 128-dimensional fusion feature vector. The first hidden layer maps the 128-dimensional vector to 64 dimensions, the second hidden layer maps the 64-dimensional vector to 128 dimensions, and the output layer maps the 128-dimensional vector to a scalar for predicting the hemoglobin content.

[0110] The working principle and beneficial effects of the above technical solution are as follows: the embodiment first constructs a blood loss prediction network and a hemoglobin prediction network including an input layer, a first hidden layer, a second hidden layer and an output layer; secondly, the fused feature, i.e., a 128-dimensional feature vector, is passed as input to a two-way fully connected network, the first hidden layer maps the 128-dimensional vector to 64 dimensions, the second hidden layer maps the 64-dimensional vector to 128 dimensions, and the output layer maps the 128-dimensional vector to a scalar for predicting blood loss; finally, the input also comes from the 128-dimensional fused feature vector, the first hidden layer maps the 128-dimensional vector to 64 dimensions, the second hidden layer maps the 64-dimensional vector to 128 dimensions, and the output layer maps the 128-dimensional vector to a scalar for predicting hemoglobin content. . Step S301 of the above scheme constructs a blood loss and hemoglobin content prediction network, designs two independent fully connected networks (one for blood loss prediction and the other for hemoglobin content prediction), each network includes an input layer, two hidden layers and an output layer; through the multi-layer nonlinear structure, the network can learn more complex mapping relationships and improve the learning ability of the model. Significance: Through independent network heads for different tasks, the learning efficiency of the model is higher, and the output of each task can be optimized specifically; separate processing can better adapt to the characteristics of each task and reduce interference between tasks. Step S302 inputs fusion features for blood loss prediction, and passes the fusion features, that is, 128-dimensional feature vectors, into the blood loss prediction network; the network extracts useful information from the input features, and after calculation, maps the 128-dimensional features to a scalar for outputting the prediction results of blood loss. Significance: By optimizing the network structure and the introduction of features, the blood loss can be predicted more accurately, and the depth of data support for clinical decision-making construction can be improved; it provides fast prediction results, can be used in clinical environments, and timely evaluates the bleeding situation of patients. Step S303 inputs the fused features to predict the hemoglobin content, and inputs the same feature vector from the fused network into the hemoglobin prediction network; a fully connected layer is also used to map the 128-dimensional features to a scalar, and the predicted hemoglobin content is output. Significance: Optimizing the output of hemoglobin content provides good support for blood loss prediction, and facilitates the model to coordinate the output of different results. Since blood loss and hemoglobin concentration are related, accurate prediction of hemoglobin content can better assess the patient's overall condition; it can quantify bleeding conditions and hemoglobin levels, provide important information to doctors, and further assist in diagnosis and treatment.

[0111] The network structure of the first fully connected network (Blood Hb Head):

[0112] Input layer: input fusion features (128 dimensions);

[0113] First hidden layer: Use the first fully connected layer (128 to 64 dimensions), usually using a nonlinear activation function (such as Sigmoid) to increase the network's expressiveness;

[0114] Second hidden layer: Use the second fully connected layer (64 to 128 dimensions) and use the activation function again to capture complex feature relationships;

[0115] Output layer: output hemoglobin content (a scalar);

[0116] The network structure of the second fully connected network (Blood Volume Head):

[0117] Input layer: also input fusion features (128 dimensions);

[0118] First hidden layer: Use the first fully connected layer (128 to 64 dimensions) and also use a nonlinear activation function (such as Sigmoid);

[0119] Second hidden layer: Use the second fully connected layer (64 to 128 dimensions) and continue to use the activation function;

[0120] Output layer: output blood loss (a scalar);

[0121] In summary, this embodiment improves the effectiveness of the model in medical image analysis. Each step forms a complete prediction process through specific network structure design, feature input and processing, and ultimately achieves accurate estimation of blood loss and hemoglobin content, providing reliable support and decision-making basis for clinical medicine. It can reasonably process complex medical data, improve the accuracy of prediction, and thus help doctors better assess the condition and make decisions.

[0122] Example 6: Figure 7 As shown, based on Embodiments 1 to 5, the model training and optimization system for blood loss estimation provided by the embodiment of the present invention includes:

[0123] The feature extraction module is responsible for obtaining the pre-made blood-soaked gauze sample and automatically learning and extracting the first type of features of the blood-soaked gauze sample through the convolutional neural network; at the same time, the color histogram features, stained area features and color moment features extracted in the HSV color space are extracted as the second type of features, which are manually designed features containing expert domain knowledge;

[0124] The feature conversion module is responsible for converting the first and second features into a unified feature space through the conversion network and completing the splicing of feature vectors;

[0125] Feature fusion module, responsible for obtaining fusion features through feature fusion network;

[0126] The regression prediction module is responsible for fusion features and outputting hemoglobin content and bleeding volume respectively through two fully connected networks.

[0127] The working principle and beneficial effects of the above technical solution are as follows: the feature extraction module of this embodiment obtains a pre-made blood-soaked gauze sample, and automatically learns and extracts the first type of features of the blood-soaked gauze sample through a convolutional neural network; at the same time, the color histogram features in the HSV color space, the red area features and the color moment features in the RGB color space are extracted as the second type of features, and the second type of features are manually designed features containing expert domain knowledge; the feature conversion module converts the first type of features and the second type of features into a unified feature space through four fully connected networks respectively, and completes the splicing of feature vectors; the feature fusion module obtains fused features through a feature fusion network; the regression prediction module outputs the fused features through two fully connected networks, respectively, and outputs the hemoglobin content and the amount of bleeding. The feature extraction module of the above scheme can capture the spatial features in the image, such as edges, colors, shapes, etc., which are usually key visual information related to the amount of bleeding and hemoglobin content by automatically learning from the convolutional neural network to extract the first type of features from the blood-soaked gauze sample; the color histogram features are extracted in the HSV color space, and the red area features and color moment features are extracted in the RGB color space. As the second type of features, the color histogram features can represent the statistics of different colors in the image, while the color moment features provide a supplement to the statistical information of the color. Significance: By combining automatic learning and manually designed features, the model can obtain more comprehensive and rich information, which helps to improve the prediction ability of the model; the second type of features contains expert domain knowledge, which can introduce the understanding of medical knowledge and provide a more rational explanation for the relationship between the amount of bleeding and the hemoglobin concentration, thereby improving the credibility of the model. The feature conversion module converts the first type of features and the second type of features through four fully connected networks respectively to ensure that they are all mapped to a unified feature space; the converted features are spliced ​​and formed into a comprehensive feature vector through a fusion network for subsequent processing. Significance: By fusing features from different sources, the model can combine complementary information to form a more discriminative feature representation, enhancing the model's understanding of complex samples; the fusion of multiple features enables the model to conduct comprehensive analysis from multiple dimensions when predicting hemoglobin concentration and blood loss, thereby improving the accuracy of prediction. The regression prediction module inputs the fused features into two independent fully connected networks; one network is responsible for outputting hemoglobin content, and the other network is responsible for outputting blood volume; each network extracts the corresponding information from the input fused features and maps it to a predicted value. Significance: Through independent output networks, the model can optimize for each specific task (such as hemoglobin content and blood loss); improve the learning efficiency and accuracy of each task; simplify the process of converting data streams into prediction results, making real-time monitoring and feedback more feasible in clinical applications, and providing doctors with timely information support in emergency situations.

[0128] In summary, through the above modules, the blood loss estimation model in this embodiment can analyze and predict the blood-soaked gauze image from different dimensions and angles. The feature acquisition module obtains important visual features, the feature conversion and fusion module combines and uniformly maps these features into a feature space, and the result output module is responsible for the specific prediction of hemoglobin content and blood loss. This not only improves the prediction ability of the model, but also achieves higher efficiency and accuracy in practical applications.

[0129] Obviously, those skilled in the art can make various changes and modifications to the present invention without departing from the spirit and scope of the present invention. Thus, if these modifications and variations of the present invention belong to the scope of equivalent technologies of the present invention, the present invention is also intended to include these modifications and variations.

Claims

1. A model training and optimization method for blood loss estimation, characterized in that: The following steps are involved: The first type of features of the samples are automatically extracted through convolutional neural networks; artificially designed features are extracted as the second type of features; The first and second features are converted to a unified feature space through a feature conversion network, the feature vectors are concatenated, and the converted features are passed through a feature fusion network to obtain fused features; The fused features are used to output the hemoglobin content and bleeding volume respectively through the regression network.

2. The model training and optimization method for blood loss estimation according to claim 1, characterized in that: The process of obtaining the first type of features includes the following steps: Obtain a sample image of blood-soaked gauze, input the sample image of blood-soaked gauze into a convolutional neural network, and obtain different feature maps by applying different convolution kernels to the image. The low-level feature map is input into the next convolution layer to obtain a feature map with a higher level of abstraction. Under the guidance of the network loss function, the convolutional network parameters are continuously adjusted through the error feedback algorithm. The feature map output by the convolutional network is the first type of feature automatically learned and extracted from the blood-soaked gauze sample.

3. The model training and optimization method for blood loss estimation according to claim 2, characterized in that: The convolutional neural network adopts the ResNet-101 structure as the convolutional neural network, which consists of 1 conv1 layer, 3 conv2_x blocks, 4 conv3_x blocks, 23 conv4_x blocks, and 3 conv5_x blocks, with a total of 1+3*(3+4+23+3)+1=101 layers.

4. The model training and optimization method for blood loss estimation according to claim 2, characterized in that: The process of generating a feature map includes the following steps: The feature maps of each layer in the convolutional network are normalized by the normalization layer, then pooled by the pooling layer, and then used as the input of the next network layer.

5. The model training and optimization method for blood loss estimation according to claim 4, characterized in that: Each convolution kernel is multiplied by the pixel value of the local area, and all the multiplication results are added together to form a single value, which represents the response of the local area to the convolution kernel.

6. The model training and optimization method for blood loss estimation according to claim 1, characterized in that: The process of obtaining the second feature includes the following steps: Obtain a sample image of blood-soaked gauze, identify the red area on the gauze, and divide the number of pixels in the red area by the total number of pixels in the image as the area feature; In RGB space, take the image mean E i , standard deviation σ i and skewness i As color moment features; 9 color moment features are obtained in three color channels; The blood-soaked gauze sample image is converted from RGB space to HSV space, and H, S, and V are quantized into multiple statistical bins respectively. For each pixel in the blood-soaked gauze sample image, its corresponding quantization level is found and counted in the corresponding statistical bin to obtain a multidimensional vector representing the number or frequency of pixels in each quantization bin.

7. The model training and optimization method for blood loss estimation according to claim 6, characterized in that: The H, S, and V channels are quantized at 12, 8, and 8 levels, respectively, and a multidimensional histogram is generated using the three quantized channels, resulting in a histogram containing 768 statistical bins.

8. The model training and optimization method for blood loss estimation according to claim 6, characterized in that: The jth pixel value of the i-th color channel is p i,j , the three color moments are expressed as follows: Where N is the number of pixels, and 9 color moment features are obtained from 3 color channels in the color space.

9. The model training and optimization method for blood loss estimation according to claim 1, characterized in that: The process of obtaining fused features through the feature fusion network includes the following steps: Obtain 768 color histogram features, 9 color moment features, 10 area features and 1024 convolution features, a total of four features; A conversion network consisting of four single-layer or multi-layer fully connected networks is used to transform the four features into a unified feature space; Through a two-layer fully connected network, the four features in the feature space are fused to obtain a 128-dimensional feature vector.

10. A model training and optimization system for blood loss estimation, characterized in that: Include: The feature extraction module is responsible for obtaining the pre-made blood-soaked gauze sample image, automatically learning and extracting the first type of features of the blood-soaked gauze sample through the convolutional neural network; extracting the artificially designed features as the second type of features, which include the blood-stained area, the color histogram features in the HSV color space, and the color moment in the RGB color space; The feature conversion module is responsible for converting the first and second features into a unified feature space through the feature conversion network and completing the splicing of feature vectors; Feature fusion module, responsible for obtaining fusion features through feature fusion network; The regression prediction module is responsible for taking the fusion features as input and outputting the hemoglobin content and bleeding volume respectively through two fully connected networks.

Images