Device and method for identifying skin condition at tracheotomy position

Abstract

The invention provides a device and method for recognizing the skin condition at the tracheotomy position, and belongs to the field of medical instruments. The device comprises an image acquisition unit, a model operation equipment unit and a display unit, the image acquisition unit is used for acquiring pictures; the model operation equipment unit is used for predicting the pictures to obtain a skin injury classification total grade score and nursing suggestions; and the display unit is used for displaying the total grade of the skin injury classification and nursing suggestions and pictures. According to the method, the accuracy and the sensitivity of identifying the damaged degree of the skin at the tracheotomy are superior to those of manual observation and evaluation, so that the postoperative care effect is greatly improved.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of medical devices, in particular to a tracheostomy skin condition recognition device and method. BACKGROUND

[0002] At present, the tracheostomy skin condition monitoring technology is relatively backward, and the method of filling the evaluation form after visual inspection is usually used, which is only based on the experience of medical staff, and different personnel observation will bring great error. Due to the difference in the evaluation of tracheostomy skin damage by different nursing workers, the damage classification and damage degree are not clear, so the nursing measures are different, and the skin damage nursing of patients is uneven.

[0003] In order to improve the accuracy of the evaluation of the skin condition of the tracheostomy of the patient and promote the recovery of the damaged skin of the tracheostomy of the patient, it is urgent to change from traditional qualitative analysis to more accurate quantitative analysis. SUMMARY

[0004] The purpose of the present application is to solve the problems existing in the prior art, and a tracheostomy skin condition recognition device and method are proposed. The device is an all-in-one machine, which uses digital image processing technology as a means, applies a neural network model, realizes the recognition of the skin picture of the tracheostomy, and makes the skin surface condition evaluation from traditional qualitative analysis to accurate quantitative analysis.

[0005] The present application is realized by the following technical solutions:

[0006] A tracheostomy skin condition recognition device, comprising: an image acquisition unit, a model running equipment unit, a display unit;

[0007] The image acquisition unit is used for acquiring pictures;

[0008] The model running equipment unit is used for predicting the pictures to obtain the total grade score of skin damage classification and nursing suggestions;

[0009] The display unit is used for displaying the total grade score of skin damage classification, nursing suggestions and pictures.

[0010] Further, the model running equipment unit comprises: an image judgment unit, an image processing unit, a model processing unit and a model post-processing unit;

[0011] The image judgment unit is used for judging the acquired pictures to obtain pictures meeting the conditions;

[0012] The image processing unit is used for processing the pictures meeting the conditions to obtain processed pictures;

[0013] The model processing unit is configured to predict the processed picture to obtain a prediction result.

[0014] The model post-processing unit is configured to calculate a total grade score of skin damage classification according to the prediction result, and obtain a care suggestion according to the total grade score of skin damage classification.

[0015] Further, the image processing unit performs data enhancement and image normalization processing on the processed picture.

[0016] Further, the model processing unit uses a deep neural network model to predict the processed picture to obtain a prediction result.

[0017] The prediction result includes skin damage classification and grade score.

[0018] Further, the deep neural network model includes a backbone network, a neck network and a head network connected in sequence.

[0019] The backbone network adopts resnet101 to extract image features.

[0020] The neck network adopts a feature pyramid network to process the extracted image features to obtain fused features.

[0021] The head network includes a classification multi-layer perceptron to obtain a prediction result according to the fused features.

[0022] Further, the head network further includes an attention mechanism layer and a concat layer.

[0023] The attention mechanism layer is configured to obtain important features of each layer of features.

[0024] The concat layer is configured to convert the important features through a concat function to obtain re-fused features.

[0025] The classification multi-layer perceptron is configured to process the re-fused features to obtain a prediction result.

[0026] Further, the deep neural network model is trained using adaptive moment estimation as an optimizer.

[0027] Further, a tracheostomy skin condition recognition method uses a tracheostomy skin condition recognition device, and the method includes:

[0028] Step 1: Collecting pictures;

[0029] Step 2: judging the pictures to obtain pictures meeting the conditions;

[0030] Step 3: Process the images that meet the criteria to obtain the processed images;

[0031] Step 4: Perform prediction on the processed image to obtain the prediction result;

[0032] Step 5: Obtain the overall classification score of skin damage and nursing recommendations based on the prediction results.

[0033] Furthermore, in a method for recognizing skin conditions at a tracheotomy site, step 4 involves using a deep neural network model to predict the processed image and obtain prediction results; the prediction results include skin lesion classification and grade score.

[0034] Furthermore, the deep neural network model includes a backbone network, a neck network, and a head network connected in sequence;

[0035] The supervisor network uses ResNet101 to extract image features;

[0036] The neck network employs a feature pyramid network to process the extracted image features and obtain fused features.

[0037] The head network includes a classification multilayer perceptron for obtaining prediction results based on the fused features.

[0038] Compared with the prior art, the beneficial effects of the present invention are: the present invention is superior to human observation and assessment in terms of accuracy and sensitivity in identifying the degree of skin damage at the tracheotomy site, thereby greatly improving the effect of postoperative care. Attached Figure Description

[0039] Figure 1 Overall technical framework diagram.

[0040] Figure 2 A simplified schematic diagram of the device casing.

[0041] Figure 3 Model structure diagram.

[0042] Figure 4 Internal processing structure diagram of Attention. Detailed Implementation

[0043] The present invention will now be described in detail with reference to the accompanying drawings:

[0044] This invention provides a device for recognizing the skin condition at the tracheotomy site. For example... Figure 1 As shown, the device of the present invention includes: an image acquisition unit, a model running device unit, and a display unit.

[0045] The image acquisition unit is used to acquire images;

[0046] The model running device unit is used to obtain a skin damage classification total grade score and a care suggestion by predicting the picture;

[0047] The display unit is used to display the skin damage classification total grade score and the care suggestion.

[0048] Further, the model running device unit comprises an image judging unit, an image processing unit, a model processing unit, and a model post-processing unit.

[0049] The image judging unit judges the picture to obtain a picture meeting the condition; the image processing unit processes the picture meeting the condition to obtain a processed picture; the model processing unit uses a deep neural network model to predict the processed picture to obtain a prediction result; and the model post-processing unit obtains a skin damage classification total grade score and a care suggestion according to the prediction result.

[0050] Further, the recognition device can further comprise a GPU processor and a large-capacity storage unit.

[0051] Preferably, the GPU processor is an NVIDIA GTX 1060, the graphics card memory is 6G, and the large-capacity storage unit is not less than 512G.

[0052] When the device is running, the image acquisition unit acquires a picture; the image judging unit judges the acquired picture to obtain a picture meeting the condition, and then sends the picture meeting the condition to the image processing unit; the image processing unit processes the picture meeting the condition to obtain a processed picture, and stores the processed picture to the large-capacity storage unit; the model processing unit calls the processed picture from the large-capacity storage unit, and predicts skin damage by using a deep neural network model to obtain a prediction result; the model post-processing unit calculates a skin damage classification total grade score according to the prediction result output by the model processing unit, matches a corresponding care suggestion, and transmits to the display unit; and the display unit displays the skin damage classification total grade score and the corresponding care suggestion.

[0053] Further, as shown in Figure 2 The recognition device further comprises a shell, the image acquisition unit and the display unit are arranged on the shell, and the model running device unit is arranged in the shell. A battery can be arranged in the shell to supply power to the image acquisition unit, the display unit, and the model running device unit. Functional keys such as a key and a photographing key can be arranged on the shell, which are all mature technologies and will not be described herein.

[0054] The following will be described in detail with examples:

[0055] The image acquisition unit is used to acquire a picture of a skin of a trachea incision part of a patient.

[0056] Specifically, the image acquisition unit adopts an existing embedded camera. When the tracheal incision of the patient is damaged, the corresponding part of the patient can be photographed and the picture can be stored in the mass storage unit.

[0057] The model running device unit: calls the picture provided by the image acquisition unit in the mass storage unit, processes and predicts the picture, and outputs the total grade of skin damage classification and nursing suggestions. Specifically, it includes:

[0058] The image determination unit: evaluates the picture collected by the image acquisition unit. Specifically, the image determination unit can use existing various image recognition indicators to evaluate the picture. The image recognition indicators can determine whether the picture meets the requirements of "clear, no occlusion, and reasonable picture brightness", etc. If it meets these requirements, it is determined that the picture is in line with the requirements, otherwise it is determined that the picture is not in line with the requirements.

[0059] The image determination unit sends the picture that meets the requirements to the image processing unit. For the picture that does not meet the requirements, the display unit prompts to retake, and the new picture after retaking replaces the original picture in the mass storage unit.

[0060] The image processing unit: used for data enhancement and image normalization processing of the picture that meets the conditions to obtain the processed picture.

[0061] Specifically, the picture is uniformly scaled to a fixed size of 320*800 pixels.

[0062] Further, data enhancement is performed, that is, data processing is performed through random flipping, cropping, translation, brightness and contrast adjustment, etc. Specifically, it includes: picture random rotation, randomly selecting an angle between 0-360 degrees for rotation, and the preferred rotation angle is between positive and negative 30 degrees; random left and right flipping; random up and down flipping; randomly selecting a value between factors 0 to 0.1 to adjust the brightness; randomly selecting a value between factors 0 to 0.1 to adjust the contrast.

[0063] Further, the picture is normalized to improve the generalization ability of the model. Specifically, the normalization processing is to convert the color image pixel 0-255 value to a value between 0-1 to obtain the processed picture.

[0064] Model processing unit: Inputs the processed image into a deep neural network model, uses the deep neural network model to make predictions, and obtains prediction results. The prediction results include: skin damage classification and grade scores. Skin damage classification includes: affected skin area, redness, erosion, blisters, and suppuration; the grade score indicates the degree of skin damage, with different scores corresponding to different degrees of skin damage. For example, each skin damage classification corresponds to four grade scores: 0, 1, 2, and 3. The lower the score, the better the skin condition around the tracheotomy, as shown in Table 1.

[0065] Table 1

[0066]

[0067]

[0068] like Figure 3 As shown, the deep neural network model is constructed using a backbone + neck + head approach, comprising a backbone network, a neck network, and a head network connected sequentially. The backbone network uses the existing ResNet101; the neck network uses the existing Feature Pyramid Network (FPN); and this invention adds an attention mechanism layer to the head network to obtain important features from each layer.

[0069] Specifically, the specific structure of the deep neural network model includes:

[0070] (1) The backbone network uses the existing ResNet101 to extract image features.

[0071] (2) The neck network uses the existing feature pyramid network (i.e., FPN) to process the image features extracted by ResNet101 to obtain pyramid features, and then performs feature fusion on the pyramid features to obtain the fused features; the feature fusion specifically involves sampling on the high-level features and transforming them through the concat function to obtain the final low-level features, i.e., the fused features.

[0072] (3) The head network processes the fused features and outputs the prediction result. The head network includes an attention mechanism layer, a concat layer (using the existing concat function), and a classification multilayer perceptron (MLP) connected in sequence. Specifically, the attention mechanism layer focuses the fused features on the important features of each layer, the concat layer transforms the important features through the concat function, that is, it performs feature fusion again to obtain the re-fused features, and finally the multilayer perceptron processes the re-fused features to obtain the prediction result (i.e., out classification).

[0073] One implementation of the constructed deep neural network model is as follows:

[0074] The spatial pyramid pooling layer of the feature pyramid network is used to process the features output by the resnet101 to obtain processed multi-scale features.

[0075] The processing utilizes the top-down and path aggregation layer in the feature pyramid network to fuse the multi-scale features in a bottom-up direction to obtain fused features. The multi-scale features include low-dimensional and high-order semantic information sampled on high-level features. The specific implementation process is as follows: the low-dimensional and high-order semantic information in the high-level features is fused through a concat function to obtain low-level features.

[0076] In this embodiment, the levels of the feature pyramid network are F1-F4, denoted as F1, F2, F3, and F4. The tensor (i.e., shape) is a 4-dimensional array, i.e., the tensor arrays are [b, 64, 80, 200], [b, 128, 40, 100], [b, 256, 20, 50], and [b, 512, 10, 25]. Here, high-level and low-level are relative, for example, F4 is a high-level feature, and F1-F3 are low-level features; F3 is a high-level feature, and F1-F2 are low-level features.

[0077]

[0078] The feature fusion from high level to low level is that the F4 features are upsampled to obtain features with the same tensor array as F3, and then processed by a concat function to obtain the final features of F3. Similarly, F2 is obtained by twice upsampling F4 to obtain features with the same tensor array as F2, and then processed by a concat function to obtain the final features of F2; F1 is obtained by thrice upsampling F4 to obtain features with the same tensor array as F1, and then processed by a concat function to obtain the final features of F1.

[0079] In this embodiment, before entering the classification multilayer perceptron of the head network, an attention mechanism is used to obtain a weight distribution, which is then applied to the features to make the model pay more attention to the features of the damaged part of the picture, so that the classification result is more accurate. After the attention mechanism processing, the tensor array is fused again by a concat function. Then, the fused features are processed by the classification multilayer perceptron to output the prediction result.

[0080] ​Specifically, the picture is subjected to backbone+FPN to obtain four layers of final features F1-F4, and then subjected to an attention mechanism layer (attention structure) to obtain four layers of features A1-A4 with attention effect. The processing principle of the attention structure is as follows:

[0081]

[0082] Among them, the roles of Q (Query) and K (Key) are to carry information between tokens, and V (Value) is the information extracted from the current token.

[0083] The processing process of the attention structure is as shown in Figure 4 .

[0084] Taking the F1 to obtain the A1 feature as an example for description:

[0085] F1 is subjected to three linear layers to obtain Q, K, and V, all of which have the same size, such as [b, 1024]. After Matmul operation, that is, dot product operation, the corresponding positions are multiplied to obtain the value of the corresponding position [b, 1024]. After scale operation, the influence of noise is eliminated. After mask and softmax operation, the result is obtained. After MatMul operation with V, the output result of attention is obtained. In this embodiment, because the softmax operation is used, the model gives a high probability value to some information that needs to be focused on, so that the feature information of the object of attention can be finally obtained.

[0086] When training the above deep neural network model, adaptive moment estimation (Adam) can be used as the optimizer.

[0087] The existing model training method can be used for training, and the process is as follows:

[0088] The pre-trained weight is loaded as the basic model weight; the optimization algorithm used is adaptive moment estimation (Adam), and the learning rate is 0.001; wherein the classification loss function uses binary cross entropy as the loss of classification and target score.

[0089] The classification loss function uses the binary cross entropy function, that is, the cross entropy loss function. The difference between two separate probability distributions P(x) and Q(x) of the same random variable can be measured by using relative entropy, that is, the divergence KL, which is defined as follows:

[0090]

[0091] Among them, x irepresents the damage classification sample point, i represents the damage classification sample number, n represents the total number of samples, D kl (p‖q) represents the discrete degree;

[0092] The smaller KL is, the closer the distribution of P(x) and Q(x) is, and the distribution of Q(x) can be approximated to P(x) by repeatedly training Q(x).

[0093] In the specific implementation, the KL discrete degree formula is further disassembled, as follows:

[0094]

[0095] Wherein, H(p(x)) represents the information entropy, H(p,q) is the cross entropy, and the KL discrete degree is equal to the cross entropy minus the information entropy.

[0096] The cross entropy formula is represented as:

[0097]

[0098] In the present application, the input picture and the picture label have been determined, so the information entropy is a constant. Since the KL discrete degree represents the difference between the real probability distribution P(x) and the predicted probability distribution Q(x), the smaller the value is, the better the prediction result is, so it is necessary to minimize the KL discrete degree, and the cross entropy is equal to the KL discrete degree plus a constant (i.e. the information entropy), so the cross entropy loss function is used to calculate the loss.

[0099]

[0100] Wherein, m represents the size of the input data set, i represents the damage classification sample, j represents the damage degree, x ij represents the value of the damage degree j in the damage classification sample i, and Loss represents the difference between the classification prediction and the actual classification, and the smaller the difference is, the more accurate the prediction is.

[0101] The model post-processing unit is used to obtain the total grade score of skin damage classification according to the prediction result output by the model processing unit, then obtain the care suggestion according to the total grade score of skin damage classification, and send the care suggestion to the display unit.

[0102] Specifically, the operation of calculating the total grade score of skin damage classification according to the prediction result output by the model processing unit includes: adding the grade scores corresponding to all skin damage classifications to obtain the total grade score of skin damage classification, for example, the grade score of damaged skin range is 1, the grade score of redness is 2, the grade score of erosion is 0, the grade score of blister is 1, and the grade score of suppuration is 0, so the total grade score of skin damage classification is 4 (1+2+0+1+0=4).

[0103] The operation of obtaining the nursing suggestion according to the total grade score includes: finding the score range where the total grade score is located, and matching the nursing suggestion corresponding to the score range. For example, the nursing suggestion corresponding to the score range of 0-5 is nursing suggestion 1, that is, gently scrub the local skin with a cotton ball dipped in normal saline, and pad the skin around the tracheostomy with sterile gauze after natural air drying. The total grade score of the skin damage classification calculated above is 4, and 4 is located in the range of 0-5, so the nursing suggestion 1 is sent to the display unit.

[0104] Display unit: used for displaying the total grade score of skin damage classification and nursing suggestion, and re-shooting prompt. The display unit uses the existing embedded display, and the resolution thereof is preferably 1080*1920, and the frequency thereof is 25HZ.

[0105] The model training method of the application is as follows:

[0106] Step 1, obtaining training samples and verification samples, the specific steps are as follows:

[0107] S1-1, obtaining pictures, and evaluating the picture quality.

[0108] The pictures of the skin of the patient at the tracheostomy site are collected by the image acquisition unit and stored in the mass storage unit.

[0109] Preferably, the image determination unit evaluates the quality of the pictures collected by the image acquisition unit by using image recognition indicators, and determines whether the pictures meet the requirements. The image recognition indicators can determine whether the pictures meet the requirements of "clear, no occlusion, and reasonable picture brightness". If these requirements are met, it is determined that the pictures meet the requirements, otherwise it is determined that the pictures do not meet the requirements. A re-shooting prompt is given for the pictures determined to be not in accordance with the requirements.

[0110] Further, the nursing staff is informed to re-shoot through the operation interface of the display unit.

[0111] Preferably, the re-shot picture replaces the old picture, and the operation process is completed through the operation interface of the display unit.

[0112] S1-2, uniform size and marking of the pictures, forming a data set.

[0113] The pictures meeting the requirements are processed by the image processing unit. Specifically, it is scaled to a fixed size, preferably the size of the verification sample picture, that is, 320*800 pixels, and is marked to obtain a marked picture, and the marked picture is stored in the mass storage unit to form a data set for subsequent model training. The data set is classified and marked by an experienced nursing team on pictures containing clinically significant findings, that is, the picture data is divided into model training set, verification set and test set according to the ratio of 7:1:2, and is marked respectively.

[0114] S1-3, data augmentation is performed on the labeled images.

[0115] The image processing unit performs data augmentation on the labeled images using rotation, flipping, same down-sampling, cropping, translation, brightness and contrast adjustment to obtain the data-augmented labeled samples.

[0116] Specifically, the labeled images are randomly rotated, a random angle between 0-360 degrees is selected for rotation, and the preferred rotation angle is between plus or minus 30 degrees; randomly left and right flip; randomly up and down flip; randomly select a value between 0 to 0.1 to adjust the brightness; randomly select a value between 0 to 0.1 to adjust the contrast.

[0117] S1-4, normalizing the data-augmented labeled samples.

[0118] Subtract the mean value of the sample from the labeled sample to achieve normalization and obtain the training samples and validation samples. Specifically, the normalization process converts the color image pixel value from 0-255 to a value between 0-1.

[0119] Step 2, use the training samples and validation samples to train the deep neural network model to obtain the trained deep neural network model.

[0120] The implementation of training the model is as follows:

[0121] Use adaptive matrix estimation (i.e. Adam) as the optimizer to train the model.

[0122] In this embodiment, the pictures with model training sample labels in the data set are input into the model, the model generates skin damage classification prediction and prediction confidence probability, and the training process is as follows:

[0123] During training, the collected data (i.e. labeled samples) is divided into training set, validation set and test set according to the ratio of 7:1:2;

[0124] The prediction effect is verified on the validation set, and as the classification loss continuously decreases, the prediction accuracy of the model also continuously increases;

[0125] The model uses Adam optimizer to optimize the loss function, which can automatically adjust the learning rate.

[0126] During the training process, the model will try to find the optimal parameter value so that the difference between the prediction result of the model on the training data set and the true result is minimized.

[0127] Specifically, the learning rate is 0.001, and after 100 iterations of the test model, the classification loss function does not decrease, so the iteration number epoch is selected as 100; the learning rate will decrease according to the cos function with the number of training times, and the decrease period is 100 rounds, and the learning rate is where i is the i-th round, and the learning rate of the last round is lr=1.57e-5.

[0128] The classification loss function uses the cross entropy function, that is, the cross entropy loss function, and the difference between two separate probability distributions P(x) and Q(x) of the same random variable can be measured by the relative entropy, that is, the KL divergence, and the formula is as follows:

[0129]

[0130] where x i represents the sample points of the damage classification sample, i represents the number of damage classification samples, n represents the total number of samples, D kl (p‖q) represents the divergence.

[0131] In the model training in this embodiment, P(x) is used to represent the real distribution of the sample, and Q(x) is used to represent the distribution predicted by the model. In this embodiment, there are 5 groups of large categories, and each large category has 4 levels of grades, and the corresponding array is Picture real distribution Predicted distribution The KL divergence is calculated as follows:

[0132]

[0133] The smaller the KL divergence, the closer the distribution of P(x) and Q(x) is, and the distribution of Q(x) can be approximated to P(x) by repeatedly training Q(x).

[0134] In a specific implementation, the KL divergence formula is decomposed as follows:

[0135]

[0136] where H(p(x)) represents the information entropy, H(p,q) is the cross entropy, and the KL divergence=cross entropy-information entropy;

[0137] The cross entropy formula is represented as:

[0138]

[0139] where H(p,q) represents the difference between the predicted result q and the real classification result p, and the smaller the value, the better the predicted result, and the closer to the actual situation.

[0140] In this embodiment, the input picture data and picture label have been determined, so the information entropy is a constant here. Since the value of KL divergence represents the difference between the true probability distribution P(x) and the predicted probability distribution Q(x), the smaller the value, the better the prediction result, so it is necessary to minimize the KL divergence, and the cross-entropy is equal to the KL divergence plus a constant (i.e. the information entropy), so the cross-entropy loss function is used to calculate the loss.

[0141]

[0142] where m represents the size of the input data set, i represents the damage classification sample, j represents the damage degree, x ij represents the value of the damage degree j in the damage classification sample i, and Loss represents the difference between the classification prediction and the actual classification.

[0143] In this embodiment, the calculation process is as follows:

[0144] If the batchsize (i.e. the size of the input data set) = 1, then

[0145] loss = -(1*log(0.4)+1*log(0.7)+1*log(0.6)+1*log(0.8)+1*log(0.9)) = 2.1123

[0146] Formula 1

[0147] In this embodiment, after iteration, the optimization is from the initial 8.17 to the final 0.002.

[0148] For example, the loss calculated in the first round is assumed to be Formula 1, and the probability of correct prediction in the first round is The maximum value of each row probability is taken as the corresponding grade classification probability of skin damage classification, i.e. [0.4, 0.7, 0.6, 0.8, 0.9], and the loss obtained is 2.1123;

[0149] In the second round, the maximum value of each row probability is taken as the corresponding grade classification probability of skin damage classification, respectively

[0150] [0.5, 0.8, 0.7, 0.81, 0.92], replace the probability in Formula 1, loss = 1.5670;

[0151] As the number of rounds increases, the probability of correct prediction also increases, and the final loss will be smaller.

[0152] The process of verifying the model is specifically: inputting the processed picture with verification sample mark into the trained model to obtain a prediction result. In one prediction embodiment of the present application, a test set with 1000 pictures is constructed, and the prediction result accuracy obtained by using the above model is 98%, which meets the requirements of clinical auxiliary use.

[0153] An implementation of obtaining a skin damage classification total grade score and nursing advice through the prediction result obtained by using the trained model is as follows:

[0154] The model post-processing unit gives a comprehensive score of skin damage according to the prediction result output by the model processing unit using the deep neural network model, that is, a skin damage classification total grade score.

[0155] In this embodiment, the prediction result is the 5 major damage classifications in Table 1, each category has a damage degree grade, and the model prediction also outputs the damage grade corresponding to the 5 major damage classifications. For example, a perfect skin, the model prediction result is The grade scores of the 5 major damage classifications are all 0, representing no, and the final result is the classification result of perfect skin. Here, the specific numerical value in the matrix can be any value close to a probability of 1, and the sum of the probabilities of each row is infinitely close to 1. The value corresponding to the maximum probability of each row is the category to which this major category belongs. For example, the 5 major damage classifications of perfect skin are all the first column category, so the probability of this column is the maximum in the row.

[0156] For example, the model prediction result is The grade scores of the 5 major damage classifications are [3, 2, 1, 2, 3], that is, the damaged skin range is more than 3 cm, moderate redness, mild erosion, 2-3 blisters, and severe suppuration. The corresponding skin damage classification total grade score is 3+2+1+2+3=11.

[0157] Further, according to the skin damage classification total grade score, the knowledge content in the mass storage unit is called to output the corresponding nursing advice matching the specified range score.

[0158] The nursing advice is knowledge formed by nursing experts with many years of rich clinical experience for different degrees of skin damage. By inputting the damage classification total grade score, the final nursing advice can be obtained.

[0159] In this embodiment, the nursing advice includes three gears:

[0160] 0-5 points: corresponding to nursing advice 1, specifically:

[0161] (1) Gently scrub the local skin with a cotton ball dipped in physiological saline;

[0162] (2) After natural air-drying, pad sterile gauze around the tracheostomy incision.

[0163] 6-10 points: corresponding to nursing suggestion 2, specifically:

[0164] (1) Wipe with a cotton ball dipped in normal saline;

[0165] (2) Iodophor disinfection;

[0166] (3) After complete drying, apply erythromycin ointment;

[0167] (4) Optionally, use adhesive soft silicone foam dressing, cut the dressing to the appropriate size according to the tracheostomy site, and cover it.

[0168] 10-15 points: corresponding to nursing suggestion 3, specifically:

[0169] (1) When the skin damage exceeds 2 hours, the wound can be considered as a contaminated wound, and when there is liquid exudation, a cotton ball dipped in normal saline can be used to clean it;

[0170] (2) Use alginate silver ion dressing and hydrocolloid dressing in combination, cut the dressing according to the size of the injury site, and cover it, ensuring that the dressing fully matches the wound, forming a sterile sealed space, and reestablishing the barrier function of the skin;

[0171] (3) If the dressing appears to be rolled, detached, or whitish, liquefied, or the patient feels uncomfortable, the wound should be re-evaluated and the doctor should be notified.

[0172] The present application also provides a tracheostomy skin condition recognition method, the method comprising:

[0173] Step 1, collect pictures;

[0174] Step 2, judge the pictures to obtain pictures meeting the conditions;

[0175] Step 3, process the pictures meeting the conditions to obtain processed pictures;

[0176] Step 4, predict the processed pictures to obtain a prediction result;

[0177] Step 5, obtain a skin damage classification total grade and nursing suggestions according to the prediction result.

[0178] The above technical solution is only one embodiment of the present application. For those skilled in the art, on the basis of the principles disclosed in the present application, various types of improvements or modifications can be easily made, and the technical solution described in the above specific embodiments is not limited to the present application, therefore the above description is only preferred, and does not have the meaning of limitation.

Claims

1. A tracheostoma skin condition recognition device, characterized in that The method comprises the following steps: An image acquisition unit, a model running device unit, and a display unit are included. The image acquisition unit is configured to acquire an image. The model running device unit is configured to predict the image to obtain a total grade score of skin damage classification and a care suggestion. The display unit is configured to display the total grade score of skin damage classification, the care suggestion, and the image.

2. The tracheostoma situation recognition device according to claim 1, characterized in that The model running device unit comprises an image judgment unit, an image processing unit, a model processing unit, and a model post-processing unit. The image judgment unit is configured to judge the acquired image to obtain an image meeting the conditions. The image processing unit is configured to process the image meeting the conditions to obtain a processed image. The model processing unit is configured to predict the processed image to obtain a prediction result. The model post-processing unit is configured to calculate the total grade score of skin damage classification based on the prediction result and obtain a care suggestion based on the total grade score of skin damage classification.

3. The tracheostoma situation recognition device according to claim 2, characterized in that The image processing unit performs data enhancement and image normalization processing on the processed image.

4. The tracheostoma skin condition recognition apparatus according to claim 2 or 3, characterized in that The model processing unit predicts the processed image by using a deep neural network model to obtain a prediction result. The prediction result comprises skin damage classification and a grade score.

5. The tracheostoma situation recognition device according to claim 4, characterized in that The deep neural network model comprises a backbone network, a neck network, and a head network connected in sequence. The backbone network adopts resnet101 and is configured to extract image features. The neck network adopts a feature pyramid network and is configured to process the extracted image features to obtain fused features. The head network comprises a classification multi-layer perceptron and is configured to obtain a prediction result based on the fused features.

6. The tracheostoma situation recognition device according to claim 5, characterized in that The head network further comprises an attention mechanism layer and a concat layer. The attention mechanism layer is configured to obtain important features of each layer of features. The concat layer is configured to convert the important features by using a concat function to obtain re-fused features. The classification multi-layer perceptron is configured to process the re-fused features to obtain a prediction result.

7. The tracheostoma situation recognition device according to claim 5, characterized in that The deep neural network model is trained by using adaptive moment estimation as an optimizer.

8. A tracheostoma skin condition recognition method characterized by, The method is implemented by using the device of any one of claims 1-7, and the method comprises the following steps: Step 1: acquiring an image. Step 2: judging the image to obtain an image meeting the conditions. Step 3: processing the image meeting the conditions to obtain a processed image. Step 4: predicting the processed image to obtain a prediction result. Step 5: obtaining a total grade score of skin damage classification and a care suggestion based on the prediction result.

9. The tracheostoma skin condition recognition method according to claim 8, characterized in that, In step 4, the processed image is predicted by using a deep neural network model to obtain a prediction result; the prediction result comprises skin damage classification and a grade score.

10. The tracheostoma skin condition recognition method according to claim 8, characterized in that, The deep neural network model comprises a backbone network, a neck network, and a head network connected in sequence. The backbone network adopts resnet101 and is configured to extract image features. The neck network adopts a feature pyramid network and is configured to process the extracted image features to obtain fused features. The head network comprises a classification multi-layer perceptron and is configured to obtain a prediction result based on the fused features.

Images