Artificial intelligence assisted peptic ulcer combined bleeding endoscopic risk assessment network and model

Abstract

The invention discloses an artificial intelligence assisted peptic ulcer combined hemorrhage endoscopic risk assessment network and model, the assessment network is a classification network, the classification network comprises more than one double-branch module, the double-branch module divides an input feature into two branches to be processed respectively, and the two branches are connected with each other. One branch extracts local features of the image through convolution operation, the other branch extracts global spatial features of the image through depth separation convolution, different normalization methods are adopted for the two branches, the convolution branch adopts batch normalization operation, and the depth separation convolution branch adopts layer normalization operation; a compression-incentive attention module is embedded in the output end of each branch, so that the classification network can adaptively enhance respective important feature channels in the two branches, and then channel splicing is performed to fuse the features of the double branches. The classification network can well perform peptic ulcer combined bleeding endoscopic image feature extraction.

Description

Technical Field

[0001] This invention belongs to the fields of medical image processing and artificial intelligence, specifically relating to an endoscopic risk assessment network and model for peptic ulcer complicated with bleeding based on artificial intelligence. Background Technology

[0002] Non-variceal upper gastrointestinal bleeding (NVUGIB) is one of the most common acute gastrointestinal conditions, with a mortality rate as high as 10%. Clinical data shows that peptic ulcer bleeding (PUB) is the leading cause of NVUGIB. In recent years, although the rate of Helicobacter pylori infection has been declining, the widespread use of nonsteroidal anti-inflammatory drugs (NSAIDs) and antiplatelet drugs has kept the prevalence of PUB at a high level.

[0003] Rebleeding in patients with penile ulceration (PUB) severely impacts prognosis, leading to prolonged hospital stays, increased mortality, and other adverse outcomes. Accurate identification of patients at high rebleeding risk is crucial for optimizing treatment plans. During endoscopic procedures, the physician's ability to quickly and accurately assess and classify risk levels directly influences treatment selection and patient outcomes. However, operator judgment is often influenced by various factors such as skill level, operator fatigue, operating environment, and endoscope performance, resulting in inconsistent outcomes and the misdiagnosis of some high-risk patients, thus missing the optimal window for treatment.

[0004] In recent years, artificial intelligence (AI) technology has been widely applied in the field of gastrointestinal endoscopy, covering multiple aspects such as colorectal adenoma detection, early cancer screening, endoscopic quality control, and bowel preparation assessment. Through training with massive amounts of endoscopic image / video data, AI can assist in improving the diagnostic accuracy and efficiency of primary care physicians. However, the current application of AI in the medical field faces three core obstacles:

[0005] 1. Black box problem: Physicians have difficulty understanding the decision-making logic of AI models (such as the basis for high-risk classification), resulting in insufficient trust and the lack of explanatory power hinders targeted optimization;

[0006] 2. Ethical and Responsibility Dilemma: When AI makes erroneous decisions, it is difficult to trace the cause and clearly define the division of responsibility;

[0007] 3. Insufficient domain adaptability: The model has poor generalization ability in different hospitals and populations.

[0008] To address obstacle 1 mentioned above, Explainable AI, or XAI, achieves transparency in the AI ​​reasoning process through technologies such as visualization, feature importance analysis, and decision path tracing. While existing research has used AI for PUB risk classification, exploration combining XAI is still in its early stages. Therefore, more innovative explainable methods are needed to break through traditional limitations and advance the integration of XAI into medical research to an advanced level. Summary of the Invention

[0009] The purpose of this invention is to provide a novel artificial intelligence-assisted endoscopic risk classification network for peptic ulcer with bleeding (PUB), and based on this network, to provide a classification model that incorporates interpretable AI. This classification model aims to assist clinicians, especially junior physicians, in reducing the risk of endoscopic misjudgment and improving the efficiency and accuracy of risk assessment for PUB patients.

[0010] The present invention provides an artificial intelligence-assisted endoscopic risk assessment network for peptic ulcer with bleeding, which is a classification network. The classification network comprises one or more dual-branch modules. Each dual-branch module divides the input features into two branches for processing. One branch uses convolution to extract local features of the image, while the other uses depthwise convolution to extract global spatial features. Different normalization methods are applied to these two branches: batch normalization for the convolution branch and layer normalization for the depthwise convolution branch. A compression-excitation attention module is embedded at the output of each branch, enabling the classification network to adaptively enhance the important feature channels in each branch. Then, the features from both branches are concatenated and fused.

[0011] The classification network also includes one or more downsampling convolutions to reduce the amount of data processed by subsequent network layers, thereby reducing the computational load. This allows the network to learn more abstract and representative image features with less computational resources, while also reducing the risk of overfitting.

[0012] Preferably, each of the dual-branch modules is preceded by a downsampling convolution.

[0013] The classification network is finally processed by global average pooling and then outputs the predicted risk level through a fully connected layer.

[0014] Another aspect of the present invention is as follows: An artificial intelligence-assisted endoscopic risk assessment model for peptic ulcer complicated with bleeding, the establishment process of which is as follows:

[0015] 1) Preprocess the endoscopic images of patients with peptic ulcer complicated with bleeding and construct an endoscopic image dataset from them. Label the endoscopic images in the endoscopic image dataset according to the constructed simplified Forrest classification system, i.e., SFCS.

[0016] In SFCS, patients' endoscopic images are divided into three categories: high risk corresponds to levels Ia and Ib in the traditional Forrest classification system, medium risk corresponds to levels IIa and IIb in the traditional Forrest classification system, and low risk corresponds to levels IIc and III in the original Forrest classification system.

[0017] 2) Divide the endoscope image dataset into different sets including training set and test set, read the endoscope images in the training set, input them into the classification network for training, test the performance of the trained model with the test set, and save the model that performs best on the test set.

[0018] The evaluation model also supports the use of the gradient-based visualization method Grad-CAM to generate heatmaps representing the contribution of the risk level predicted by the model.

[0019] International guidelines recommend using endoscopic prognostic scoring systems to assess the risk of NVUGIB patients, with the Rockall score, Glasgow-Blatchford score (GBS), and Forrest classification system being widely used. However, the Rockall score and GBS have limited clinical application due to their complex calculation processes. Since its introduction in 1974, the Forrest classification system has become the mainstream tool for assessing the risk of rebleeding in poop ulcers worldwide. This system classifies ulcers into six grades based on their endoscopic appearance: spurting bleeding (Forrest Ia), oozing bleeding (Forrest Ib), exposed vessels (Forrest IIa), and adherent blood clots (Forrest IIb). Additionally, there are black-based ulcers (Forrest IIc) and clean-based ulcers (Forrest III), and treatment plans are recommended according to the grade.

[0020] However, this system has significant shortcomings in real-world clinical scenarios:

[0021] Classification ambiguity: The boundaries between categories are unclear, relying heavily on the subjectivity of the rater. In the real world, there have been instances of Forrest IIc being misclassified as Forrest IIa, and IIa being misclassified as III. A study in Minnesota showed that nearly 25% of PUB patients had classification inconsistencies.

[0022] Treatment decision bias: Inconsistent classification can lead to overtreatment or undertreatment. For example, a study by Yidan Lu et al. found that only 64.5% of high-risk patients underwent endoscopic hemostasis, while 9.8% of low-risk patients were over-intervened. Therefore, the traditional Forrest classification system urgently needs optimization to improve its clinical applicability.

[0023] Based on the similarity of treatment plans, this invention merges the 6-level classification into three categories: high, medium, and low risk. As verified (see the detailed implementation section), this can save judgment time, improve judgment efficiency, and enhance the consistency of judgments by the assessors.

[0024] To address the persistent issues of unclear boundaries between categories and reliance on the subjectivity of raters, this invention constructs a novel classification network structure. By combining it with the gradient-based visualization method Grad-CAM, interpretable AI technology is used to clearly define images. After training, this model demonstrates accuracy comparable to traditional Forrest grading by humans (senior resident physicians and above) in assessing the risk of rebleeding in PUB patients. Furthermore, by displaying a heatmap influencing the decision-making process, physicians can easily understand the model's decision logic and use it to assist in assessing the risk of rebleeding in PUB patients. This not only improves assessment efficiency but also reduces the risk of misjudgment caused by physicians' lack of proficiency, fatigue, environmental factors, etc., thereby enhancing the accuracy of the assessment.

[0025] The Grad-CAM in this invention not only matches well with the classification network of this invention, but also has low computational cost.

[0026] Step 2) The metrics used to evaluate the performance of the model are: model confusion matrix, accuracy, sensitivity, and specificity.

[0027] The preprocessing in step 1) includes image cropping and image enhancement, wherein the image cropping is used to reduce background areas with interference.

[0028] Beneficial effects:

[0029] This invention, based on the simplified Forrest classification system (SFCS) and combined with innovative interpretable AI technology, establishes an AI-assisted endoscopic risk assessment model for peptic ulcer bleeding (PUB), namely XAI-SFCS. This model can assist clinicians in reducing the risk of endoscopic misjudgment and improving the efficiency and accuracy of risk assessment for PUB patients. This invention pioneers a new method for combining interpretable AI with endoscopic technology in the field of digestive endoscopy, solving the "black box" problem of existing AI models used for PUB risk assessment, and laying the foundation for the development of subsequent endoscopic AI diagnostic technology. Attached Figure Description

[0030] Figure 1 This embodiment reflects the complete technical approach of the artificial intelligence-assisted endoscopic risk assessment model for peptic ulcer complicated with bleeding;

[0031] Figure 2 The study analyzed the dispersion of the results from 20 evaluations conducted by 20 physicians using both the traditional Forrest grading system and the SFCS grading system.

[0032] Figure 3 This is a schematic diagram of the XAI-SFCS network structure;

[0033] Figure 4 A schematic diagram of a heatmap generated using a gradient-based visualization method;

[0034] Figure 5a , 5b 5c represents the views of the confusion matrix on the test set, internal validation set, and external validation set, respectively.

[0035] Figure 6 A comparison of the time taken for manual grading using traditional Forrest and corresponding image analysis using the XAI-SFCS system. Detailed Implementation

[0036] The artificial intelligence-assisted endoscopic risk assessment network and model for peptic ulcer complicated with bleeding (PUB) of the present invention will be further described in detail below with reference to the embodiments and accompanying drawings.

[0037] Figure 1 This reflects the complete technical approach of the AI-assisted endoscopic risk assessment model for peptic ulcer complicated with bleeding in this embodiment. For example... Figure 1 As shown, firstly, endoscopic images and related clinical data of PUB patients from Sun Yat-sen Memorial Hospital (Shenzhen-Shantou Central Hospital) and five branch centers of Sun Yat-sen University were acquired to establish and validate the Simplified Forrest Classification System (SFCS). Secondly, the XAI-SFCS was established using endoscopic images of PUB patients from the five branch centers. Finally, the XAI-SFCS was clinically validated using data from Shenzhen-Shantou Central Hospital. Figure 1 The validation operation in the context of using an external validation set.

[0038] The creation process of the above model is as follows:

[0039] I. Data Collection Standards

[0040] Image selection criteria:

[0041] 1. Patients aged 18 years or older but under 75 years old;

[0042] 2. Exclude contraindications for endoscopy;

[0043] 3. Patients diagnosed with peptic ulcer complicated by bleeding who underwent emergency endoscopy at Sun Yat-sen Memorial Hospital of Sun Yat-sen University (Shenzhen-Shantou Central Hospital) and related branch centers. Diagnostic criteria for active bleeding include: black or bloody stools, hematemesis, decreased hemoglobin, hemodynamic instability, elevated uric acid nitrogen, decreased urine output, and active bleeding observed endoscopically.

[0044] 4. The endoscopic image is clear (without excessive mucus, bubbles, bile, food residue, blurriness, darkness, defocus, fuzziness, or halo).

[0045] 5. Complete clinical data.

[0046] Image exclusion criteria:

[0047] 1. Applicants must be under 18 years of age or over 75 years of age;

[0048] 2. Individuals with multiple underlying medical conditions who are unable to undergo endoscopic evaluation;

[0049] 3. Individuals with mental disorders who are unable to undergo endoscopic examination;

[0050] 4. Regardless of whether it is the first or second bleeding episode, death due to excessive blood loss makes it impossible to trace the observer;

[0051] 5. Malignant ulcer;

[0052] 6. Poor image quality (e.g., excessive mucus, bubbles, bile, food residue, blurriness, darkness, defocus, fuzziness, or halo) and incomplete clinical data.

[0053] Informed consent: This is a retrospective study, therefore informed consent is waived.

[0054] II. Establishing SFCS

[0055] Based on clinical needs, we have established a simplified Forrest classification system, namely the Simplified Forrest Classification System (SFCS). This system divides PUB patients into three categories: high risk (corresponding to levels Ia and Ib of the traditional Forrest classification system, active bleeding), intermediate risk (corresponding to levels IIa and IIb of the traditional Forrest classification system, exposed vessels and blood clots, inactive bleeding), and low risk (corresponding to levels IIc and III of the traditional Forrest classification system, a black background indicates that the hemoglobin has been oxidized after bleeding, and a clean background in category III indicates no recent bleeding).

[0056] Treatment decisions for Ia and Ib are similar, as are those for IIa and IIb, and those for IIc and III. Therefore, merging them into three categories—high, medium, and low risk—can save judgment time, improve judgment efficiency, and enhance the consistency of judgment among assessors.

[0057] III. Verification of SFCS

[0058] First, two senior physicians and 18 senior resident physicians (with more than three years of experience in endoscopy) from five branch hospitals were asked to rate the selected 1458 (main center) + 10151 (branch center) endoscopic images using the traditional Forrest grading system. Figure 1 As shown, these physicians were then trained in recognizing, memorizing, and outputting the new SFCS system. To avoid memory retention, they were asked to apply the SFCS grading system to rate the aforementioned endoscopic images after a one-month interval.

[0059] After 20 physicians evaluated each image using both the traditional Forrest grading system and the SFCS grading system, the degree of dispersion within each of the 20 evaluations is as follows: Figure 2 As shown. For the traditional Forrest grading system, the standard deviation among 20 results from 20 physicians grading the same image was 0.22905; for the SFCS grading system, the standard deviation among 20 results from 20 physicians grading the same image using this system was 0.05494. This indicates that the dispersion of results (lower dispersion indicates higher consistency) differs among physicians using different systems. A paired t-test comparing the differences in standard deviations between the two systems showed a p-value < 0.05, indicating a significant difference in inter-rater consistency between the two systems. Figure 2 ).

[0060] Comparing the mean standard error of the results from the two evaluation systems, it was found that the SFCS (mean standard error of 0.00055) had lower data dispersion than the traditional Forrest grading system (mean standard error of 0.00227) (P<0.0001). In other words, the SFCS improved the consistency among evaluators, while the traditional Forrest grading system resulted in greater differences among evaluators (Figure 2).

[0061] Simultaneously, the 7-day rebleeding rate was assessed in patients who provided images. Rebleeding was defined as the recurrence of any of the following within 7 days: melena or hematochezia, hematemesis, decreased hemoglobin, hemodynamic instability, or confirmation of a conventional Forrest classification of Ia-Ib via endoscopy. Results showed that the 7-day rebleeding rate was 70.3% for conventional Forrest classification Ia-b and 79.2% for SFCS high-risk patients (P>0.05); 40.9% for conventional Forrest classification IIa-b and 77.4% for SFCS intermediate-risk patients (P<0.05); and 46.5% for conventional Forrest classification IIc-III and 18.6% for SFCS low-risk patients (P<0.05). It is evident that SFCS is more sensitive in assessing high and medium risk, more specific in assessing low risk, and more meaningful in guiding treatment.

[0062] IV. Establishment and Validation of XAI-SFCS

[0063] 10,151 endoscopic images of real-world PUB patients from 5 branch hospitals were preprocessed to construct an endoscopic image dataset. Image preprocessing employed image enhancement methods such as cropping and rotation. Since the original endoscopic images contained large areas of black background, which typically contained auxiliary text information such as examination time and equipment details, a appropriately sized rectangle was used to crop the images to reduce interference from the background for feature extraction. The cropped images were then rotated to increase data diversity and improve the generalization ability of the classification network. The preprocessed endoscopic images formed the basis of the endoscopic image dataset.

[0064] Next, the endoscopic image dataset was randomly divided into training, test, and validation sets. 80% of the total number of endoscopic images in each category (high, medium, and low in SFCS) was allocated to the training set, with the remaining 10% allocated to the test and validation sets respectively. Endoscopic images from the training set were read and input into the classification network for training. A set number of training epochs was used to test the model's performance on the test set, and the best-performing model on the test set, XAI-SFCS, was saved. Model evaluation in this embodiment included the classification model's confusion matrix, accuracy, sensitivity, and specificity.

[0065] The classification network structure used in this embodiment is as follows: Figure 3As shown, the network consists of multiple (four shown in the figure) bi-branch modules, downsampling convolutions, global average pooling layers, and fully connected layers. Downsampling convolutions are placed before each bi-branch module to reduce the amount of data processed by subsequent network layers, thereby reducing computational cost. This allows the network to learn more abstract and representative image features with fewer computational resources, while also reducing the risk of overfitting. The number of bi-branch modules is related to computational resources; too few modules make it difficult to cover multi-scale lesions, while too many modules may cause the loss of features from small targets.

[0066] In this embodiment, the structure of the dual-branch module is as follows: Figure 3 As shown, the input features are divided into two branches for processing: one uses convolution to extract local features of the image, and the other uses depthwise segregating convolution to extract global spatial features. Different normalization methods are used based on the different characteristics of these two branches: batch normalization is used for the convolution branch, and layer normalization is used for the depthwise segregating convolution branch. To enable the classification network to adaptively enhance the important feature channels in each branch, a compression-encouragement attention module is embedded in each branch to adjust the channel weights of the features in both branches. Then, channel concatenation is performed to fuse the information from the two branches. Finally, after global average pooling, the predicted risk level is output through a fully connected layer.

[0067] The dual-branch module of the classification network in this embodiment can effectively extract features, ensuring the classification accuracy of the network. Furthermore, this classification network employs a lightweight network structure, reducing the number of parameters and computational load during training by using local convolutions and depthwise separable convolutions in parallel.

[0068] To enhance the interpretability of the classification model, a gradient-based visualization method, Grad-CAM, is used to generate a heatmap representing the contribution of the model to the predicted risk level. Specifically, gradient backflow is performed based on the predicted risk level to calculate the gradient information of the target class score with respect to the feature map of the last convolutional layer. Then, global average pooling is performed on the gradient tensor to generate weights for each channel in the feature map. The weights are then weighted and summed with the feature map, and finally activated by the ReLU function to obtain the final heatmap, as shown below. Figure 4 As shown.

[0069] Compared to other commonly used methods, such as CAM and LIME, the Grad-CAM method has the following advantages:

[0070] 1. Grad-CAM has low computational cost; 2. Adding it can maintain model performance well without modifying the structure of the classification network or retraining; 3. It can intuitively locate the regions that play a key role in classification decisions, thus providing interpretability.

[0071] CAM requires modification of the network structure, and LIME suffers from high computational costs. This embodiment, Grad-CAM, offers significant advantages in terms of computational overhead, model performance, and intuitive localization.

[0072] The expressions for the evaluation metrics Accuracy, Sensitivity, and Specificity in this embodiment are as follows:

[0073] The accuracy rate can be calculated using formula (1):

[0074] (1)

[0075] Sensitivity can be calculated using formula (2):

[0076] (2)

[0077] Specificity can be calculated using formula (3):

[0078] (3)

[0079] Among them, TP (True Positive) represents a true positive, TN (True Negative) represents a true negative, FP (False Positive) represents a false positive, and FN (False Negative) represents a false negative.

[0080] The confusion matrices of the classification model XAI-SFCS in this embodiment on the test set, internal validation set, and external validation set are as follows: Figure 5a , 5b As shown in Figures 5c, the comparison of classification evaluation metrics (accuracy, sensitivity, and specificity) on the test set, internal validation set, and external validation set is shown in Table 1.

[0081] Table 1 Comparison of classification evaluation metrics on the test set, internal and external validation sets.

[0082]

[0083] V. Verification of the diagnostic efficiency of XAI-SFCS

[0084] The diagnostic time of the XAI-SFCS classification model in this embodiment for each image was statistically analyzed. The diagnostic time of 20 physicians using the Forrest score for each corresponding image was compared. Finally, a paired t-test was performed, and the results are as follows: Figure 6As shown, the average time required for manual diagnosis using the Forrest grading system is 3.2079 minutes, while the average time required for diagnosis using the XAI-SFCS classification model of this invention is 0.1803 minutes, demonstrating the superior advantage of the XAI-SFCS classification model.

[0085] VI. Technical Advantages of the Invention

[0086] This invention, based on the simplified Forrest classification system (SFCS) and combined with innovative interpretable AI technology, establishes an AI-assisted endoscopic risk assessment model for peptic ulcer bleeding, called XAI-SFCS. This lays the foundation for the development of subsequent endoscopic AI diagnostic technologies, and its advantages are as follows:

[0087] 1) Optimization of grading standards and improvement of clinical applicability

[0088] Clinical trials have shown that optimizing the grading system enhances the operability of the grading standards, which helps reduce treatment decision-making bias caused by ambiguity in classification and lowers the risk of overtreatment and undertreatment for patients.

[0089] (ii) Evaluation effectiveness of AI-assisted SFCS

[0090] The artificial intelligence assessment model based on SFCS demonstrated accuracy comparable to traditional Forrest classification by human (senior resident physicians and above) in assessing the risk of rebleeding in PUB patients. Figure 5a While SFCS (Sustainable Surveillance System) has advantages such as reduced time and increased assessment consistency, it can save manpower and reduce labor costs. The model effectively identifies high-risk bleeding signs (such as projectile bleeding and exposed vessels) by extracting features from endoscopic images, achieving a technological leap from subjective experience-based judgment to objective quantitative assessment.

[0091] (iii) Consistency and Transparency of Risk Assessment Empowered by XAI Technology

[0092] By introducing interpretable artificial intelligence (XAI) technology, the SFCS assessment system maintains the same risk discrimination capability as the traditional Forrest classification. It also makes the assessment process traceable by visualizing the decision path (such as annotating key vascular features that affect the classification in the image) and analyzing the importance of features (such as quantifying the contribution of ulcer diameter to the high-risk classification). This feature not only solves the "black box" problem of traditional AI, but also further improves the consistency of classification results by aligning the decision-making logic of clinicians with that of the AI ​​system.

[0093] (iv) A dual breakthrough in diagnostic efficiency and standardization

[0094] In the clinical evaluation of PUB endoscopic images, the AI-assisted SFCS system demonstrates significant efficiency advantages: compared to the manual diagnostic process based on the traditional Forrest classification, the system reduces the evaluation time per image by 94.4%, while improving the consistency of classification among physicians of different seniority by 76%. This efficient and standardized evaluation model is particularly suitable for the triage of large numbers of patients and the priority intervention of critically ill patients in emergency scenarios.

Claims

1. An artificial intelligence assisted endoscopic risk assessment network for peptic ulcer with hemorrhage, which is a classification network, characterized in that, The classification network comprises one or more double-branch modules, which separate the input features into two branches for processing, one using convolution operation to extract local features of the image, and the other using depthwise separable convolution to extract global spatial features of the image, and different normalization methods are used for the two branches, batch normalization operation is used for the convolution branch, and layer normalization operation is used for the depthwise separable convolution branch, a compression-excitation attention module is embedded at the output end of each branch, so that the classification network can adaptively enhance the important feature channels in each branch, and then the features of the two branches are spliced and fused.

2. The evaluation network of claim 1, wherein, The classification network further comprises one or more down-sampling convolutions.

3. The evaluation network of claim 2, wherein, Each of the double-branch modules is provided with one down-sampling convolution.

4. The evaluation network of claim 3, wherein, The classification network finally outputs the predicted risk level through a fully connected layer after global average pooling.

5. An artificial intelligence assisted endoscopic risk assessment model for peptic ulcer with bleeding, characterized in that, The establishment process is as follows: 1) The endoscopic images of patients with peptic ulcer combined with bleeding are preprocessed to construct an endoscopic image dataset, and the endoscopic images in the endoscopic image dataset are labeled according to the constructed simplified Forrest classification system (SFCS); In the SFCS, the endoscopic images of patients are divided into three categories, wherein the high-risk corresponds to the levels Ia and Ib in the traditional Forrest classification system, the medium-risk corresponds to the levels IIa and IIb in the traditional Forrest classification system, and the low-risk corresponds to the levels IIc and III in the original Forrest classification system; 2) The endoscopic image dataset is divided to form different sets including a training set and a test set, the endoscopic images in the training set are read and input into the classification network of any one of claims 1-4 for training, the performance of the trained model is tested by using the test set, and the model with the best performance on the test set is saved; The evaluation model also supports using a gradient-based visualization method Grad-CAM to generate a heat map representing the contribution degree of the model to the predicted risk level.

6. The model of claim 5, wherein, The indicators used to evaluate the performance of the model in step 2) are the confusion matrix, accuracy, sensitivity and specificity of the model.

7. The model of claim 5, wherein, The preprocessing in step 1) includes image cropping and image enhancement, and the image cropping is used to reduce the background area with interference.

Images