Fluorescent dermatophyte detection method based on pseudo label and lightweight model

Abstract

The invention discloses a fluorescent dermatophyte detection method based on a pseudo label and a lightweight model. The method comprises the following steps: S1, collecting a dermatophyte picture data set of fluorescent staining of a patient; s2, generating a label for the data set based on a false label technology of SAM, performing iterative correction, performing data preprocessing, and dividing the data into a training set, a test set and a verification set; s3, constructing a lightweight target detection network, sending training set data into the lightweight target detection network for multiple rounds of training, obtaining an optimal lightweight model through the verification set, and accurately identifying a fungus image on the test set and outputting a detection result; and S4, carrying out statistical analysis on the fungus category and quantity of a detection result, and judging whether a patient suffers from dermatophyte related diseases or not. According to the invention, the time cost and workload of manual labeling are reduced. According to the invention, the detection speed is obviously improved, and the hardware demand and cost are reduced. And the generalization ability of the model in various fluorescent dyeing scenes is enhanced.

Description

Technical Field

[0001] The present invention belongs to the technical field of artificial intelligence and medical image detection, and specifically relates to a fluorescent skin fungus detection method based on pseudo-labels and lightweight models. Background Art

[0002] Skin fungal infection is a stubborn disease that poses a serious threat to human health. The commonly used fungal detection method is fungal fluorescent staining microscopy, but the traditional detection method mainly relies on manual observation of images under the microscope. This method has problems such as strong subjectivity and low efficiency. It is highly dependent on the experience and judgment of the inspectors and is easily affected by visual fatigue, leading to the risk of misdiagnosis or missed diagnosis. Therefore, in order to improve the efficiency and accuracy of detection, efficient and automated deep learning image detection technology has gradually become the focus of research. Deep learning methods can automatically extract features from a large number of fluorescent staining images, perform high-precision classification and positioning, and greatly reduce the interference of human factors.

[0003] However, current neural networks often require a large number of high-quality real labels for training. These labels not only require a lot of manpower and time, but also require professionals to ensure the accuracy and consistency of the labels. Very time-consuming and labor-intensive, they are all common deep learning target detection models. How to efficiently generate accurate labels and reduce the burden of data labeling is an important challenge in the development of neural network models. Commonly used neural networks, such as convolutional neural networks (CNNs) and Transformer-based structural networks, although they perform well in processing complex image tasks, generally face the problems of large computational workload and high resource consumption. These models are usually composed of a large number of ordinary convolutional layers, complex encoder-decoder structures, and multi-layer feature extraction modules, and have high requirements for hardware computing power. On PCs or high-performance servers, computing resources are sufficient to meet model requirements. However, on low-computing power or mobile devices, limited hardware resources are difficult to support the high computing load of the model, which seriously affects detection efficiency and performance. Therefore, designing a lightweight network structure to significantly reduce computing costs and memory usage while ensuring accuracy, so as to achieve rapid detection on resource-constrained devices, is an urgent problem to be solved.

[0004] Therefore, it is necessary to design a new method for fungal detection. Summary of the invention

[0005] The purpose of the present invention is to provide a fluorescent skin fungus detection method based on pseudo-labels and lightweight models, so as to solve the problem mentioned in the background technology that when the current skin fungal infection is detected by deep learning image detection technology, a large number of high-quality real labels are often required for training. These labels not only require a lot of manpower and time, but also require professionals, which is very time-consuming and labor-intensive.

[0006] To achieve the above object, the present invention provides a fluorescent skin fungus detection method based on pseudo-labels and lightweight models, comprising the following steps: S1. Collect a dataset of fluorescently stained skin fungi images from patients; S2, based on the pseudo-labeling technology of SAM, generates labels for the data set and iteratively modifies them. After data preprocessing, the data is divided into training set, test set and validation set; S3. Build a lightweight target detection network, send the preprocessed training set data into the lightweight target detection network for multiple rounds of training, and obtain the best lightweight model through the validation set, which can accurately identify fungal images on the test set and output the detection results; S4. Conduct statistical analysis on the fungal types and quantities in the test results to determine whether the patient has skin fungus-related diseases.

[0007] In a specific embodiment, in step S1, the characteristics based on which the data is collected include: age group, gender, health status, immunosuppression, whether or not the data suffer from skin diseases, and geographical location.

[0008] In a specific implementation, the step S2 specifically includes the following steps: S2.1, generate initial annotations, and use the SAM network to automatically generate initial pseudo labels based on the input fluorescent skin fungus images; S2.2, by introducing experts to verify and adjust the initial labels generated by SAM one by one, optimize and iterate the labels; S2.3, the set threshold standard is used to evaluate the quality of the labels after each round of optimization by the F1 score to ensure the reliability of the pseudo-labels. If the pseudo-labels do not meet the set threshold standard, the labeling optimization and iterative correction are performed again until the accuracy of the labels meets the requirements; S2.4. Preprocess the data set, and then divide the preprocessed data set into a training set, a validation set, and a test set.

[0009] In a specific implementation, in step S2.4, the preprocessing includes image normalization, size adjustment, denoising and contrast enhancement.

[0010] In a specific implementation, the lightweight target detection network includes a feature extraction network, a neck network, and a head network.

[0011] In a specific implementation, the feature extraction network includes a C1 layer, a C2 layer, a C3 layer, and a C4 layer arranged in sequence; data passes through the C1 layer, the C2 layer, the C3 layer, and the C4 layer in sequence; the C1 layer includes a Conv module for an initial convolution operation; the C2 layer, the C3 layer, and the C4 layer all include a GhostConv module and a GhostC3 module, and all pass through the GhostConv module first and then through the GhostC3 module.

[0012] In a specific embodiment, the neck network includes a P2 layer, a P3 layer and a P4 layer; the output end of the C4 layer is connected to the input end of the P4 layer, and the P4 layer includes a GhostC3 module and a Concat module, which first passes through the Concat module and then passes through the GhostC3 module; the P2 layer and the P3 layer both include an Upsampling module, a GhostC3 module and a Concat module, and both first pass through the Upsampling module, then the Concat module, and then the GhostC3 module. The Concat module of the P3 layer also inputs data from the output of the C3 layer, and the Concat module of the P2 layer also inputs data from the output of the C2 layer.

[0013] In a specific embodiment, the head network converts the input feature map into the required number of output channels through a convolution operation, including the classification result (Cls), bounding box regression (Reg), and target confidence (Obj).

[0014] Compared with the prior art, the present invention has the following beneficial effects: Since fluorescent skin images have a large amount of data and contain a large number of annotated instances, manual annotation is time-consuming and laborious. Based on pseudo-labeling technology, the present invention significantly reduces the time cost and workload of manual annotation. In addition, traditional detection methods are difficult to balance between accuracy and efficiency, and it is difficult to achieve fast operation on low-computing power devices. The present invention significantly improves the detection speed and reduces hardware requirements and costs through lightweight design. At the same time, the optimized network architecture enhances the generalization ability of the model in a variety of fluorescent staining scenarios.

[0015] The present invention proposes a pseudo-labeling technology based on SAM, which can greatly reduce manpower and material resources and quickly generate high-quality labels through processes such as initial annotation generation, expert iterative correction, and quality assessment. And the present invention continues to propose a lightweight target detection network structure, and the overall design consists of three parts: feature extraction network, neck network and head network. The feature extraction network retains rich semantic information and reduces computational complexity by alternately stacking GhostConv and GhostC3 modules. The neck network uses a multi-scale feature fusion strategy to effectively integrate features at different levels to enhance the detection accuracy of the model. The head network only contains one detection head, which uses the fused features for classification and regression, simplifying the network structure.

[0016] While maintaining high detection accuracy, the present invention significantly reduces the model size and computing requirements, and greatly improves the detection speed.

[0017] In addition to the above-described objects, features and advantages, the present invention has other objects, features and advantages. The present invention is further described in detail below. BRIEF DESCRIPTION OF THE DRAWINGS

[0018] The drawings constituting a part of this application are used to provide a further understanding of the present invention. The exemplary embodiments of the present invention and their descriptions are used to explain the present invention and do not constitute an improper limitation of the present invention. In the drawings: Figure 1 is a flow chart of an embodiment of the present invention; Figure 2 is a schematic diagram of a lightweight target detection network according to an embodiment of the present invention; Figure 3 It is a schematic diagram of a fluorescent sample of skin fungi according to an embodiment of the present invention. DETAILED DESCRIPTION

[0019] The embodiments of the present invention are described in detail below. The specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0020] Example 1 A fluorescent skin fungus detection method based on pseudo-labels and lightweight models of the present invention comprises the following steps: S1. Collect a dataset of fluorescently stained skin fungi images from patients.

[0021] In step S1, the features collected include: age group, gender, health status, immunosuppression, skin disease, and geographical location. The purpose of collecting fluorescent skin fungus image data from people of different age groups, genders, and other diverse characteristics is to build a comprehensive and representative data set.

[0022] S2, based on the pseudo-labeling technology of SAM, generates labels for the data set and iteratively modifies them. After data preprocessing, the data is divided into training set, test set and validation set.

[0023] The step S2 specifically includes the following steps: S2.1. Perform initial annotation generation, and use the SAM network to automatically generate initial pseudo-labels based on the input fluorescent skin fungus image; that is, automatically and preliminarily determine the category of fungi in the fluorescent skin fungus image.

[0024] S2.2. The initial labels generated by SAM are verified and adjusted one by one by introducing experts to optimize and iteratively correct the labels; that is, the fungal categories of the labels are verified and adjusted by manual verification.

[0025] S2.3, the set threshold standard is used to evaluate the quality of the labels after each round of optimization by the F1 score to ensure the reliability of the pseudo-labels. If the pseudo-labels do not meet the set threshold standard, the labeling optimization and iterative correction are performed again until the accuracy of the labels meets the requirements; S2.4. Preprocess the data set, and then divide the preprocessed data set into a training set, a validation set, and a test set.

[0026] In step S2.4, the preprocessing includes image normalization, size adjustment, denoising and contrast enhancement.

[0027] S3. Build a lightweight target detection network, send the preprocessed training set data into the lightweight target detection network for multiple rounds of training, and obtain the best lightweight model through the validation set, which can accurately identify fungal images on the test set and output the detection results; The lightweight target detection network includes a feature extraction network, a neck network and a head network.

[0028] In the lightweight object detection network, the GhostConv module is used to generate redundant features through cheap operations, reducing the number of parameters and calculations. The GhostC3 module can further improve the ability to extract multi-level features by stacking multiple GhostConv modules. GhostConv and GhostC3 form a lightweight feature extraction network, neck network, and head network.

[0029] In the lightweight feature extraction network part, GhostConv and GhostC3 are overlapped and combined, which retains rich semantic expression capabilities and provides efficient and high-quality feature maps for subsequent target detection tasks. The feature extraction network includes C1 layer, C2 layer, C3 layer and C4 layer arranged in sequence; data passes through C1 layer, C2 layer, C3 layer and C4 layer in sequence; C1 layer includes Conv module for initial convolution operation; C2 layer, C3 layer and C4 layer all include GhostConv module and GhostC3 module, and all pass through GhostConv module first and then through GhostC3 module.

[0030] The input image size is 640×640×3. After the initial convolution operation (Conv) in the C1 layer, the output feature map size becomes 320×320×64. In the C2 layer, the GhostConv module is used to further extract features, and the number of channels is increased to 320×320×128. After being processed by the GhostC3 module, the output is 320×320×256. In the C3 layer, the feature map size is reduced to 160×160×512 through the GhostConv operation and remains unchanged in the GhostC3 module. Finally, in the C4 layer, the feature map is reduced to 80×80×1024 through GhostConv, and further optimized through GhostC3, and the final output is 80×80×1024.

[0031] In the lightweight neck network part, a feature pyramid structure is constructed, and top-down feature fusion and lateral connection are performed. Combined with Upsampling and Concat operations, high-level features are upsampled and concatenated (Concat) with feature maps from the lower layer, so that the network can effectively detect objects at different scales to retain more detailed information and improve detection accuracy.

[0032] The neck network includes a P2 layer, a P3 layer and a P4 layer; the output end of the C4 layer is connected to the input end of the P4 layer, and the P4 layer includes a GhostC3 module and a Concat module, which first passes through the Concat module and then through the GhostC3 module; the P2 layer and the P3 layer both include an Upsampling module, a GhostC3 module and a Concat module, and both first pass through the Upsampling module, then the Concat module, and then the GhostC3 module. The Concat module of the P3 layer also inputs data from the output of the C3 layer, and the Concat module of the P2 layer also inputs data from the output of the C2 layer.

[0033] The 80×80×1024 feature map output by the C4 layer is directly passed to the P4 layer in the neck network. This feature map is upsampled and fused with the 160×160×512 feature map generated by the C3 layer to form the P3 layer feature map. Then, the P3 layer feature map is upsampled and fused with the 320×320×256 feature map from the C2 layer to obtain the P2 layer feature map.

[0034] In the lightweight head network part, only one head is used to detect objects of different sizes. The P2 feature map is directly passed to the head network, which converts the input feature map into the required number of output channels through convolution operations, including classification results (Cls), bounding box regression (Reg), and target confidence (Obj).

[0035] S4. Conduct statistical analysis on the fungal types and quantities in the test results to determine whether the patient has skin fungus-related diseases.

[0036] The experimental results of the lightweight model of the present invention are compared with those of the existing method, and the results are shown in Table 1: Table 1 Comparison of experimental results of the lightweight model of the present invention and the existing method As shown in Table 1, the evaluation indicators Params (M), FLOPs (G), mAP 50 、mAP 50-95 They represent the number of model parameters, the number of floating-point operations, the average accuracy at a threshold of 50%, and the average accuracy in a threshold range of 50% to 95%. The method of the present invention significantly reduces the number of parameters and computational complexity, and performs well in multiple accuracy evaluation indicators. Compared with traditional models, it achieves a smaller model size and faster reasoning speed while maintaining a high detection accuracy.

[0037] The above contents are further detailed descriptions of the present invention in combination with specific preferred embodiments, and it cannot be determined that the specific implementation of the present invention is limited to these descriptions. For ordinary technicians in the technical field to which the present invention belongs, several simple deductions and substitutions can be made without departing from the concept of the present invention, which should be regarded as belonging to the protection scope of the present invention.

Claims

1. A fluorescent skin fungus detection method based on pseudo-labels and lightweight models, characterized in that: The following steps are involved: S1. Collect a dataset of fluorescently stained skin fungi images from patients; S2, based on the pseudo-labeling technology of SAM, generates labels for the data set and iteratively modifies them. After data preprocessing, the data is divided into training set, test set and validation set; S3. Build a lightweight target detection network, send the preprocessed training set data into the lightweight target detection network for multiple rounds of training, and obtain the best lightweight model through the validation set, which can accurately identify fungal images on the test set and output the detection results; S4. Conduct statistical analysis on the fungal types and quantities in the test results to determine whether the patient has skin fungus-related diseases.

2. The fluorescent skin fungus detection method based on pseudo-labels and lightweight models according to claim 1 is characterized in that: In step S1, the characteristics based on which the data is collected include: age group, gender, health status, immunosuppression, whether the patient suffers from skin diseases, and geographical location.

3. The fluorescent skin fungus detection method based on pseudo-labels and lightweight models according to claim 1 is characterized in that: The step S2 specifically includes the following steps: S2.1, generate initial annotations, and use the SAM network to automatically generate initial pseudo labels based on the input fluorescent skin fungus images; S2.2, by introducing experts to verify and adjust the initial labels generated by SAM one by one, optimize and iterate the labels; S2.3, the set threshold standard is used to evaluate the quality of the labels after each round of optimization by the F1 score to ensure the reliability of the pseudo-labels. If the pseudo-labels do not meet the set threshold standard, the labeling optimization and iterative correction are performed again until the accuracy of the labels meets the requirements; S2.

4. Preprocess the data set, and then divide the preprocessed data set into a training set, a validation set, and a test set.

4. The fluorescent skin fungus detection method based on pseudo-labels and lightweight models according to claim 3 is characterized in that: In step S2.4, the preprocessing includes image normalization, size adjustment, denoising and contrast enhancement.

5. The fluorescent skin fungus detection method based on pseudo-labels and lightweight models according to claim 1 is characterized in that: The lightweight target detection network includes a feature extraction network, a neck network and a head network.

6. The fluorescent skin fungus detection method based on pseudo-labels and lightweight models according to claim 5 is characterized in that: The feature extraction network includes a C1 layer, a C2 layer, a C3 layer and a C4 layer arranged in sequence; data passes through the C1 layer, the C2 layer, the C3 layer and the C4 layer in sequence; the C1 layer includes a Conv module for an initial convolution operation; the C2 layer, the C3 layer and the C4 layer all include a GhostConv module and a GhostC3 module, and all pass through the GhostConv module first and then through the GhostC3 module.

7. The fluorescent skin fungus detection method based on pseudo-labels and lightweight models according to claim 6 is characterized in that: The neck network includes a P2 layer, a P3 layer and a P4 layer; the output end of the C4 layer is connected to the input end of the P4 layer, and the P4 layer includes a GhostC3 module and a Concat module, which first passes through the Concat module and then through the GhostC3 module; the P2 layer and the P3 layer both include an Upsampling module, a GhostC3 module and a Concat module, and both first pass through the Upsampling module, then the Concat module, and then the GhostC3 module. The Concat module of the P3 layer also inputs data from the output of the C3 layer, and the Concat module of the P2 layer also inputs data from the output of the C2 layer.

8. The fluorescent skin fungus detection method based on pseudo-labels and lightweight models according to claim 5 is characterized in that: The head network converts the input feature map into the required number of output channels through convolution operations, including classification results, bounding box regression, and target confidence.

Images