A leucorrhea fluorescence image detection method integrating frequency domain feature enhancement and multi-task artifact decoupling

Abstract

The application discloses a leucorrhea fluorescence image detection method integrating frequency domain feature enhancement and multi-task artifact decoupling, and belongs to the technical field of medical test image processing. The method first performs adaptive preprocessing according to red and blue channel statistics, then generates a candidate label through frequency domain amplitude learnable modulation that preserves phase information, then realizes mycelium and scratch artifact decoupling by using a multi-task detection network containing an independent scratch branch and combining with an orthogonal constraint loss, and finally performs closed-loop relearning on low-confidence and conflict samples. Based on a desensitization clinical data set verification, the mAP@0.5 (average precision mean) of the application on the test set is 96.3%0.6%. The application can be used for automatic counting and auxiliary interpretation in a test device.

Description

Technical Field

[0001] This invention relates to the fields of medical laboratory image processing and artificial intelligence technology, specifically to a method for detecting leukorrhea fluorescence images that integrates frequency domain feature enhancement and multi-task artifact decoupling, which can be used for automatic counting and auxiliary interpretation in laboratory equipment. Background Technology

[0002] In clinical laboratory testing, leukorrhea fluorescence imaging is characterized by large sample sizes, diverse target types, and complex background noise. Traditional manual microscopic examination suffers from low efficiency, significant subjective variability, and insufficient repeatability. Existing automated detection schemes mainly suffer from: insufficient generalization due to background differences between red and blue zones; false alarms caused by confusion between slide scratches and fungal hyphae; missed detections due to weak edges of small targets; and high cost of full manual annotation, making iteration difficult. Therefore, it is necessary to propose a technical solution that balances accuracy, artifact resistance, and continuous learning capabilities. Summary of the Invention

[0003] This invention discloses a method for detecting leukorrhea fluorescence images that integrates frequency domain feature enhancement and multi-task artifact decoupling. The proposed method includes the following steps:

[0004] S1. Obtain a white band fluorescence field image, calculate the red-blue partition discrimination index, and execute the red zone processing procedure or the blue zone processing procedure according to the index.

[0005] S2. Input the preprocessed image into the frequency domain enhancement Transformer pre-annotation module to obtain candidate annotations;

[0006] S3. A multi-task artifact decoupling detection model is trained based on candidate annotations and manual review results. The detection model includes 11 types of cell / microbial targets and 1 type of quality control microsphere target; scratches are detected as an independent artifact category.

[0007] S4. Using the trained detection model, output the counting results of 12 types of medical targets and the scratch artifact filtering results of the image under test.

[0008] S5. Add low-confidence samples and class conflict samples to the pool and perform incremental retraining to update the model parameters.

[0009] The partitioning criterion in step S1 is:

[0010]

[0011] In the formula, The red channel mean. This is the average value of the blue channel. To prevent zero constant, This is an indicator for distinguishing between red and blue zones.

[0012] Preferably, in step S1: when When the red zone processing procedure is executed, The blue zone processing procedure is executed at that time, and the threshold is... The range of values ​​is .

[0013] Preferably, the red area processing flow includes at least local contrast limiting adaptive equalization and bilateral filtering, and the blue area processing flow includes at least gamma correction and background suppression.

[0014] Preferably, in step S2, the frequency domain enhancement Transformer pre-annotation module satisfies:

[0015]

[0016]

[0017]

[0018] In the formula, For the input feature map, It is a two-dimensional discrete Fourier transform. It is a complex feature map in the frequency domain. It is an amplitude spectrum. It is the phase spectrum. The learnable weight matrix has a size that is similar to... Similarly, random initialization is performed using the Xavier method, and the model is updated along with the Xavier method using gradient descent. It is Hadamard element-wise multiplication. It is the Sigmoid function. It is the modulation coefficient, and its value range is... . It is to learn the modulated amplitude spectrum. It is a two-dimensional discrete Fourier inverse transform. It is to take the real part of the complex number. It is the frequency domain enhanced feature map obtained from the reconstruction.

[0019] Preferably, the enhanced features and spatial features are fused using gated fusion:

[0020]

[0021] In the formula, For spatial domain characteristics, These are trainable parameters, with initial values ​​of And the range of values ​​is , This is the fused feature map.

[0022] Preferably, the total loss function of the detection model in step S3 is:

[0023]

[0024] in, It is the main detection branch loss, used to optimize the detection performance of 12 target categories. It's damage from scratches. It is orthogonal constraint loss. It is edge consistency loss. This is the loss weighting coefficient, here. , , ,

[0025]

[0026] For classifying losses, For bounding box regression loss, The target confidence loss.

[0027] Preferably, the orthogonal constraint loss for:

[0028]

[0029] In the formula, For the sample size, and The first The scratch feature vector and hyphae feature vector of each sample were extracted from the global average pooling layer of the medical target detection head and the scratch detection head, respectively, and both were 256-dimensional feature vectors. Represents the dot product of vectors; express norm, To prevent zero constant.

[0030] Preferably, the edge consistency loss for:

[0031]

[0032] In the formula, For the first Predicted edge map for each sample, For the corresponding labeled edge map, The gradient operator can be represented by the first-order gradient calculation method, such as the Sobel operator, Prewitt operator, or Roberts operator. express Norm.

[0033] Preferably, the scratch loss for:

[0034]

[0035]

[0036]

[0037]

[0038] In the formula, Classify the loss as scratches. For the regression loss of the scratch box, Loss of confidence in the target; The total number of candidate boxes. The number of positive samples The set of positive samples; For scratch category labels, To predict the probability of a scratch; and These are the predicted bounding box and the true bounding box, respectively. The area of ​​intersection. The area of ​​the union; For target-oriented tags, This is the predicted confidence level for the target.

[0039] Preferably, the 12 categories of medical targets include: leukocytes, fungal hyphae, pseudohyphae, epithelial cells, clue cells, fungal spores, trichomonas, Gardnerella vaginalis, Staphylococcus aureus, Streptococcus, Lactobacillus, and quality control microspheres; scratches are detected as an independent artifact category.

[0040] Preferably, the sample inclusion rule in step S5 is: the detection confidence level is less than... Or category conflict score greater than The incremental retraining cycle is sky.

[0041] Preferably, the detection system of the present invention includes: a partitioning processing module configured to execute step S1; a frequency domain enhancement Transformer pre-labeling module configured to execute step S2; a multi-task artifact decoupling detection module configured to execute steps S3 and S4; and a closed-loop relearning module configured to execute step S5.

[0042] The beneficial effects of this invention are as follows: This invention reduces the impact of distribution offset and false alarms from hyphae-scratch patterns through joint optimization of partition preprocessing, frequency-domain learnable modulation, and artifact decoupling, thereby improving the detection capability of weak edge targets; on the desensitization clinical test set, mAP@0.5 96.3% 0.6%, which has engineering application value. Attached Figure Description

[0043] Figure 1 The system's overall flowchart is shown in the diagram.

[0044] 101: Image Acquisition and Desensitization;

[0045] 102: Partition Adaptive Preprocessing;

[0046] 103: Frequency Domain Enhanced Transformer Pre-annotation;

[0047] 104: Manual review;

[0048] 105: Multi-task artifact decoupling training;

[0049] 106: Model Deployment;

[0050] 107: Online reasoning and output of counting targets in 12 categories;

[0051] 108: Low-confidence / conflicting samples are included in the pool;

[0052] 109: Periodic review and incremental training.

[0053] Figure 2 The flowchart for red-blue partitioning is shown in the figure.

[0054] 201: Input fluorescence image;

[0055] 202: Calculate channel statistics ;

[0056] 203: Calculate the discriminant index threshold ;

[0057] 204: Judgment: ? ;

[0058] 205: Red zone processing flow: CLAHE + bilateral filtering;

[0059] 206: Blue area processing workflow: Gamma correction + background suppression;

[0060] 207: Scale normalization and intensity normalization;

[0061] 208: Output preprocessed image.

[0062] Figure 3 The diagram shows the pre-annotated structure of the frequency domain enhanced Transformer.

[0063] 301: Input Features ;

[0064] 302: Transformation;

[0065] 303: Amplitude-phase decomposition yields the amplitude spectrum. and phase spectrum ;

[0066] 304: Amplitude can be learned modulation. ;

[0067] 305: Frequency domain reconstruction, ;

[0068] 306: Reconstruction yields frequency domain enhancement features. ;

[0069] 307: Spatial Feature Branch, Extracting Spatial Domain Features ;

[0070] 308: Gating Fusion ;

[0071] 309: Transformer self-attention encoding;

[0072] 310: Candidate box / mask output.

[0073] Figure 4 This is a structural diagram of a multi-task artifact decoupling detection model. (See diagram for details.)

[0074] 401: Input preprocessed image;

[0075] 402: Shared feature extraction module (Backbone + Neck);

[0076] 403: Medical target detection head, outputting 12 types of cell / microbial targets;

[0077] 404: Scratch detection head, outputs scratch artifact detection results;

[0078] 405: Hyphae Feature Vector ;

[0079] 406: Scratch Feature Vector ;

[0080] 407: Orthogonal constraint loss , used to decouple hyphae from scratch features;

[0081] 408: Total Loss Function ,in, , , ;

[0082] 409: Inference Output: Count results for 12 categories of medical targets;

[0083] 410: Inference output: Scratch artifact filtering results.

[0084] Figure 5 The flowchart for closed-loop relearning is shown in the figure:

[0085] 501: Online inference results;

[0086] 502: Confidence level and conflict assessment;

[0087] 503: Judgment: or ;

[0088] For each candidate target, the model outputs a class probability vector, where each element represents the confidence level that the target belongs to a certain class. Let this vector be sorted in descending order of probability values. , ,…

[0089]

[0090] and These are the first and second largest class confidence scores for the same candidate target, arranged in descending order in the class probability vector;

[0091] 504: Sample inclusion in the pool (low confidence / conflicting sample library);

[0092] 505: Expert review and revision annotation;

[0093] 506: Incremental training;

[0094] 507: New model version released;

[0095] 508: Redeploy online system. Detailed Implementation

[0096] This invention adopts a combined technical approach of "partition adaptive preprocessing + frequency domain enhanced pre-labeling + multi-task artifact decoupling detection + closed-loop relearning".

[0097] 1. Partition Adaptive Preprocessing

[0098] The partition type is determined by the following formula:

[0099]

[0100] In the formula, The red channel mean. This is the average value of the blue channel. To prevent zero constant, This is an indicator for distinguishing between red and blue zones.

[0101] 2. Frequency Domain Enhanced Transformer Pre-annotation

[0102] Perform frequency domain transformation on the input features:

[0103]

[0104] While maintaining the phase, learnable modulation of the amplitude is performed:

[0105]

[0106] And reconstruct and enhance features:

[0107]

[0108] Enhanced features and spatial features are fused together as follows:

[0109]

[0110] In the formula, For spatial domain characteristics, These are trainable parameters, with initial values ​​of And the range of values ​​is , This is the fused feature map.

[0111] 3. Multi-task artifact decoupling detection

[0112] The detection network consists of a main detection branch and independent scratch branches, and the total loss function is:

[0113]

[0114] in:

[0115]

[0116]

[0117]

[0118]

[0119]

[0120]

[0121]

[0122] Used to suppress the coupling between scratch features and hyphal features.

[0123] 4. Closed-loop relearning

[0124] Low-confidence samples and conflicting samples are automatically added to the pool, periodically manually reviewed, and incrementally trained to continuously optimize the model.

[0125] Example 1: Dataset and Annotation

[0126] A total of 18,624 desensitized clinical fluorescence visual field images were used, with a resolution of [resolution missing]. ;according to The dataset was divided into training, validation, and test sets. The annotation categories included 12 medical targets plus 1 artifact (scratches), employing a "two-person annotation + expert arbitration" process.

[0127] Example 2: Partition Adaptive Preprocessing

[0128] calculate:

[0129]

[0130] when (For example Perform contrast-limited adaptive histogram equalization (CLAHE) and bilateral filtering; when Perform gamma correction and background suppression. Gamma correction can be expressed as:

[0131]

[0132] in, For input pixel intensity, To correct the pixel intensity, This is the intensity normalization constant; when the image is 8-bit grayscale, . Gamma coefficient, Enhance the shadows. Suppress highlights. The range of values ​​is .

[0133] Example 3: Frequency Domain Enhancement Pre-labeling

[0134] Input features Follow these steps to process:

[0135]

[0136]

[0137]

[0138]

[0139] Will Input the Transformer self-attention encoding module, which is used to process the fused feature map. Perform self-attention encoding to generate candidate labels.

[0140] Example 4: Multi-task artifact decoupling training

[0141] Training input size is The batch size is 16, the training duration is 300 epochs, the optimizer is AdamW, and the initial learning rate is... .use:

[0142]

[0143] Inference threshold: confidence level NMS threshold .

[0144] Example 5: Closed-loop relearning

[0145] The inclusion condition is the predicted confidence level. Or category conflict score ,in:

[0146]

[0147] in and These represent the first and second largest class confidence scores for the same candidate target, arranged in descending order in the class probability vector.

[0148] Every A manual review and incremental update are performed once a day.

[0149] Example 6: Performance Results

[0150] On the desensitization clinical test set, this invention achieved mAP@0.5=96.3%. mAP@0.5:0.95=69.8%, 0.6% 97.1%.

[0151] The above embodiments are only used to illustrate the present invention and are not intended to limit the present invention; any equivalent substitutions or modifications made without departing from the spirit and substance of the present invention shall fall within the protection scope of the present invention.

Claims

1. A method for detecting leukorrhea fluorescence images that integrates frequency domain feature enhancement and multi-task artifact decoupling, characterized in that, Includes the following steps: S1. Obtain a white band fluorescence field image, calculate the red-blue partition discrimination index, and execute the red zone processing procedure or the blue zone processing procedure according to the index. S2. Input the preprocessed image into the frequency domain enhancement Transformer pre-annotation module to obtain candidate annotations; S3. A multi-task artifact decoupling detection model is trained based on candidate annotations and manual review results. The detection model includes 11 types of cell / microbial targets and 1 type of quality control microsphere target; scratches are detected as an independent artifact category. S4. Using the trained detection model, output the counting results of 12 types of medical targets and the scratch artifact filtering results of the image under test. S5. Add low-confidence samples and class conflict samples to the pool and perform incremental retraining to update the model parameters. The partitioning criterion in step S1 is: In the formula, The red channel mean. This is the average value of the blue channel. To prevent the value of a zero constant, the range of values ​​is: Preferred , This is an indicator for distinguishing between red and blue zones.

2. The method according to claim 1, characterized in that, In step S1: when When the red zone processing procedure is executed, The blue zone processing procedure is executed at that time, and the threshold is... The range of values ​​is .

3. The method according to claim 1, characterized in that, In step S1: the red area processing flow includes at least local contrast limiting adaptive equalization and bilateral filtering, and the blue area processing flow includes at least gamma correction and background suppression.

4. The method according to claim 1, characterized in that, In step S2, the frequency domain enhancement Transformer pre-annotation module satisfies the following: In the formula, Input feature map; It is a two-dimensional discrete Fourier transform; It is a complex feature map in the frequency domain; It is the amplitude spectrum; It is the phase spectrum; The weight matrix is ​​a learnable weight matrix; It is Hadamard element-wise multiplication; It is the Sigmoid function; It is the modulation coefficient, and its value range is... ; It involves learning the modulated amplitude spectrum; It is a two-dimensional discrete Fourier inverse transform; It takes the real part of the complex number; It is the frequency domain enhanced feature map obtained from the reconstruction.

5. The method according to claim 4, characterized in that, Enhanced features and spatial features are fused using gating: In the formula, Features of the spatial domain These are trainable parameters, with initial values ​​of And the range of values ​​is , This is the fused feature map.

6. The method according to claim 1, characterized in that, The total loss function of the detection model in step S3 is: in, It is the main detection branch loss, used to optimize the detection performance of 12 target categories; It's damage from scratches. It is orthogonal constraint loss. It is edge consistency loss; This is the loss weighting coefficient, here. , , , For classifying losses, For bounding box regression loss, The target confidence loss.

7. The method according to claim 6, characterized in that, The orthogonal constraint loss for: In the formula, For the sample size, and The first Scratch feature vector and hyphal feature vector of each sample; Represents the dot product of vectors; express Norm, To prevent zero constant; The edge consistency loss for: In the formula, For the first Predicted edge map of each sample, For the corresponding labeled edge map, The gradient operator can be represented by the first-order gradient calculation method, such as the Sobel operator, Prewitt operator, or Roberts operator. express Norm; The scratches for: In the formula, Classify the loss as scratches. For the regression loss of the scratch box, Loss of confidence in the target; The total number of candidate boxes. The number of positive samples The set of positive samples; For scratch category labels, To predict the probability of a scratch; and These are the predicted bounding box and the true bounding box, respectively. The area of ​​intersection. The area of ​​the union; For target-oriented tags, This is the predicted confidence level for the target.

8. The method according to claim 1, characterized in that, The 12 categories of medical targets include: leukocytes, fungal hyphae, pseudohyphae, epithelial cells, clue cells, fungal spores, trichomonas, Gardnerella vaginalis, Staphylococcus aureus, Streptococcus, Lactobacillus, and quality control microspheres; scratches are detected as a separate artifact category.

9. The method according to claim 1, characterized in that, The sample inclusion rule in step S5 is: the prediction confidence is less than... Or category conflict score greater than The incremental retraining cycle is sky.

10. A system for detecting leukorrhea fluorescence images that integrates frequency domain feature enhancement and multi-task artifact decoupling, characterized in that, include: The partitioning processing module is configured to execute step S1 as described in claim 1; the frequency domain enhancement Transformer pre-labeling module is configured to execute step S2; the multi-task artifact decoupling detection module is configured to execute steps S3 and S4; and the closed-loop relearning module is configured to execute step S5.

Images