A semi-supervised segmentation method and system based on amplitude-aware deep reinforcement learning

Abstract

This invention discloses a semi-supervised segmentation method and system based on amplitude-aware deep reinforcement learning, belonging to the field of image processing technology. The method includes acquiring scalp HR-MR images as a dataset and labeling the image segmentation targets; obtaining synthetic images through a deep reinforcement learning strategy; applying perturbation to the input unlabeled image, and jointly inputting the synthetic and original images into an encoder to extract image features step-by-step, training a semi-supervised image segmentation network; processing low-order features of the unlabeled image through a phase alignment strategy; inputting the high-order features of the perturbated unlabeled image into a cross-power spectrum correlation constraint module; adding weights to calculate the total loss, updating network parameters through backpropagation, and optimizing the image segmentation network. The method described in this invention enables the model to adaptively learn the effects of different perturbations and layer structure changes, achieving good segmentation even with extremely thin layers, avoiding noise interference, and providing more global information that greatly aids in segmentation.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, specifically to a semi-supervised segmentation method and system based on amplitude-aware deep reinforcement learning. Background Technology

[0002] Androgenetic alopecia (AGA) is a common hair loss condition that impacts quality of life and mental health. Accurate segmentation of the scalp tissue layers is crucial for understanding the mechanisms and staging of AGA. High-resolution magnetic resonance imaging (HR-MR) is a highly effective tool for staging and assessment. Deep learning-based segmentation methods have achieved significant success by training on large-scale labeled datasets. However, accurate segmentation remains challenging due to the lack of large-scale labeled datasets, low image quality, and the effects of tissue deformation, which negatively impact the generalization performance of segmentation models.

[0003] To address these issues and improve model generalization ability, some semi-supervised methods focus on utilizing unlabeled data. These can generally be divided into two categories: pseudo-label methods and consistency regularization methods. Most existing methods tend to emphasize the utilization of unlabeled data while neglecting the importance of effective data augmentation. Most methods use CutMix as the default data augmentation strategy. While these methods can alleviate overfitting and enhance data diversity to some extent, their randomness and non-adaptive nature make them difficult to guarantee effectiveness in complex situations and sometimes even prone to side effects. Secondly, for the use of unlabeled data, pseudo-label methods typically generate pseudo-labels by pre-training on labeled data. In many cases, this improvement is often limited. Furthermore, consistency regularization can be applied to multiple aspects such as the model, data, and task. These methods aim to enhance the robustness of the model and reduce its sensitivity to noise. However, various types of strong and weak perturbation consistency and the design of multiple encoders or decoders ignore the fact that the distribution of random noise differs significantly from the distribution of real noise. Summary of the Invention

[0004] In view of the above-mentioned problems, the present invention is proposed.

[0005] Therefore, the technical problems solved by this invention are: the scarcity of high-quality labeled data in existing segmentation tasks, the complexity of the organizational structure, the difficulty of thin-layer segmentation, and the impact of noise on segmentation accuracy.

[0006] To address the aforementioned technical problems, this invention provides the following technical solution: a semi-supervised segmentation method based on amplitude-aware deep reinforcement learning, comprising:

[0007] We collected scalp HR-MR images as a dataset and labeled the image segmentation targets.

[0008] The input labeled and unlabeled images are decomposed to obtain style and structural information. Through a deep reinforcement learning strategy, the style features of the labeled image with high confidence labels and the unlabeled image are adaptively swapped to obtain the synthetic image.

[0009] The input unlabeled image is perturbed to obtain weakly perturbed and strongly perturbed images. The synthesized image and the original image are then jointly input into the encoder to extract image features step by step. The features of the labeled image and the unlabeled image are jointly input into the decoder and classifier to train a semi-supervised image segmentation network.

[0010] Weakly perturbed images, strongly perturbed images, and synthetic images of unlabeled images are input into the encoder to obtain low-order features, and the low-order features of the unlabeled images are processed through a phase alignment strategy.

[0011] The high-level features of the perturbed unlabeled image are input into the cross-power spectrum correlation constraint module to obtain constraints on the segmentation results; weights are added to calculate the total loss, and the network parameters are updated through backpropagation to optimize the image segmentation network.

[0012] As a preferred embodiment of the semi-supervised segmentation method based on amplitude-aware deep reinforcement learning described in this invention, the step of labeling the data segmentation targets includes having professionals label the data segmentation targets and dividing the data into training and testing sets; wherein the training set contains labeled data and unlabeled data, and the testing set contains only unlabeled data;

[0013] The image decomposition includes processing the labeled image. and unlabeled images The labeled images were obtained by performing frequency domain decomposition using Fourier transform. The amplitude is Phase is Unlabeled images The amplitude is .

[0014] As a preferred embodiment of the semi-supervised segmentation method based on amplitude-aware deep reinforcement learning described in this invention, the deep reinforcement learning policy includes a policy network P in the deep reinforcement learning model that, according to the policy... Generate an action , This represents the action of transferring style from an unlabeled image to a labeled image, where For the action space, For state space; The state represents the input received by the model at time t. ;

[0015] Mark the amplitude of the image The data is fed into a deep reinforcement learning model, where each pixel is treated as an agent. The value network V in the deep reinforcement learning model evaluates the expected reward v based on the state after the current action, while the actual reward r is obtained from environmental feedback. The agent iterates by adjusting its actions according to the advantage function, dynamically learning a scaling factor. ,use Control the size of the switching area;

[0016] Calculate the side length W of the exchange region on the amplitude maps of the labeled and unlabeled images, and then... The central W×W region is assigned Corresponding area, and then The images are then fused and style-transferred to obtain the synthesized image through inverse Fourier transform. ;

[0017] Will The data is fed into a segmentation network to obtain the segmentation results. The task of the value network V is to analyze the current prediction results of the model. To evaluate the expected reward value that the agent's actions can obtain; to evaluate the predicted segmentation results and real labels Calculate the cross-entropy loss and use it as the true reward r to update the value network V.

[0018] As a preferred embodiment of the semi-supervised segmentation method based on amplitude-aware deep reinforcement learning described in this invention, the training of the semi-supervised image segmentation network includes constructing a value network loss. Strategy network loss ;Calculate the loss of the synthesized image using cross-entropy , will lose The segmentation network is used as a reward value to evaluate the style transfer enhancement of the image. The ability to segment and calculate value loss. ;Calculate the total loss of the reinforcement learning model;

[0019] Weakly perturbed images of unlabeled images and strongly disturbed images The encoder is used to obtain features from weakly and strongly perturbated images, and low-order features from the weakly and strongly perturbated images are extracted respectively. and Then and Input phase alignment module.

[0020] As a preferred embodiment of the semi-supervised segmentation method based on amplitude-aware deep reinforcement learning described in this invention, the phase alignment strategy includes obtaining the phase alignment through Fourier transform. The phase spectrum is , The phase spectrum is Original image The phase spectrum is ;

[0021] Low-order features of weak and strong perturbation images and The filtered features are fed into the multi-scale INC module, and the phase spectrum is obtained by applying Fourier transform. , Calculate the phase consistency loss between images with weak and strong perturbations. Align images with weak and strong perturbations.

[0022] As a preferred embodiment of the semi-supervised segmentation method based on amplitude-aware deep reinforcement learning described in this invention, the cross-power spectrum correlation constraint module includes processing the low-level features of weakly and strongly enhanced unlabeled images input into the CPSC module.

[0023] The perturbed unlabeled image Advanced features The input is a Cross-Power Spectrum Correlation Constraint (CPSC) module. This CPSC module uses two classifiers with different initializations to predict the segmentation results of the input features, resulting in two logical outputs. and The prediction results are subjected to Fourier transform. The normalized cross power spectrum is calculated for the frequency domain representation of the two different prediction results. The real part of the spectrum is then taken to generate a spatial domain correlation graph to capture global constraints. The spatial offset of the two prediction results is calculated and the offset is gradually corrected to align the two prediction results, thereby obtaining the correlation constraints on the segmentation results. The correlation constraints on the segmentation results are applied to the output of the original decoder to obtain the constrained prediction results and construct consistency constraints.

[0024] As a preferred embodiment of the semi-supervised segmentation method based on amplitude-aware deep reinforcement learning described in this invention, wherein: updating network parameters through backpropagation includes, for Perform an inverse Fourier transform to convert to the spatial domain, take the real part, and generate a spatial domain correlation graph. :

[0025] Find the two peaks in the correlation graph and calculate the spatial offset. ;

[0026] Using spatial offset Align to After alignment Constructing a related graph , representing the probabilistic mean of the two segmentation results;

[0027] Using related graphs Adjust the decoder's logic output to obtain the constrained logic output; apply a softmax operation to the constrained logic output to obtain the prediction result. This makes the decoder's prediction Consistency alignment is performed with the constrained predictions, and the relevant loss is calculated; KL divergence is used for the unlabeled images with the two perturbations to reduce the gap between the predictions.

[0028] Simultaneously with the feature processing of the CPSC module, the corresponding high-order and low-order features are fused and fed into the decoder. Multi-scale convolution is used to reduce the dimensionality of the high-order features to be consistent with the low-order features. The two are then concatenated using a cat operation. After concatenation, the image is passed through a classifier to obtain a prediction. Bilinear interpolation is used to enlarge the image to the same size as the original image. The prediction result is obtained, and the total loss is calculated.

[0029] As a preferred embodiment of the semi-supervised segmentation system based on amplitude-aware deep reinforcement learning described in this invention, it includes an ADDA module, a phase alignment PHA module, and a cross-power spectrum correlation CPSC module.

[0030] The ADDA module is used to combine fast Fourier transform and inverse Fourier transform to achieve adaptive image style transfer.

[0031] The phase alignment (PHA) module is used to enable the model to focus only on the structural information of the image at the feature level through a phase consistency strategy.

[0032] The Cross Power Spectrum Correlation (CPSC) module is used to capture global constraints through cross power spectrum calculation, thereby optimizing prediction alignment and consistency in layer segmentation tasks.

[0033] A computer device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement a semi-supervised segmentation method based on amplitude-aware deep reinforcement learning.

[0034] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of a semi-supervised segmentation method based on amplitude-aware deep reinforcement learning.

[0035] The beneficial effects of this invention are as follows: Deep reinforcement learning is used to extract style information from a large amount of unlabeled data, enabling the model to adaptively learn the effects of different perturbations and layer structure changes, thus allowing the model to segment even extremely thin layers well. Fourier transform is used to separate the phase and amplitude of features, constructing a phase consistency loss, which allows the model to focus only on structural information, thereby better avoiding noise interference. Feature constraints constructed using cross-power spectra can effectively enable the model to acquire more global information, and since layer structure features often span the entire image, more global information greatly aids in segmentation. Attached Figure Description

[0036] To more clearly illustrate the technical solutions of the embodiments of the present invention, the drawings used in the following description of the embodiments will be briefly introduced. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0037] Figure 1 The first embodiment of the present invention provides an overall flowchart of a semi-supervised segmentation method based on amplitude-aware deep reinforcement learning.

[0038] Figure 2 The following is a flowchart illustrating the invention process of a semi-supervised segmentation method based on amplitude-aware deep reinforcement learning, provided for the second embodiment of the present invention.

[0039] Figure 3 The image shows the segmentation results of a scalp HR-MR dataset compared with other methods, based on a semi-supervised segmentation method using amplitude-aware deep reinforcement learning, as provided in the second embodiment of the present invention.

[0040] Figure 4 The image shows the segmentation results of a semi-supervised segmentation method based on amplitude-aware deep reinforcement learning, provided in the second embodiment of the present invention, compared with other methods on a retinal dataset. Detailed Implementation

[0041] To make the above-mentioned objects, features, and advantages of the present invention more apparent and understandable, specific embodiments of the present invention will be described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not all of them. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort should fall within the protection scope of the present invention.

[0042] Example 1, referring to Figure 1 As an embodiment of the present invention, a semi-supervised segmentation method based on amplitude-aware deep reinforcement learning is provided, comprising:

[0043] S1: Obtain HR-MR images of the scalp as a dataset and label the image segmentation targets.

[0044] Furthermore, the annotation of the data segmentation target includes having professionals annotate the data segmentation target to divide the data into a training set and a test set; wherein both the training set and the test set contain labeled data and unlabeled data.

[0045] Obtain N+M HR-MR images as a dataset, and select N images from them as the labeled dataset:

[0046]

[0047] Among them, the label For image Corresponding tags; Indicates the sequence number. M images form an unlabeled dataset:

[0048]

[0049] Crops all images to a uniform size.

[0050] It should be noted that the image decomposition includes decomposing the labeled image. and unlabeled images The labeled images were obtained by performing frequency domain decomposition using Fourier transform. The amplitude is The phase of the labeled image is Unlabeled images The amplitude is .

[0051] S2: Decompose the input labeled and unlabeled images to obtain style and structural information. Through a deep reinforcement learning strategy, adaptively exchange the style features of the labeled image with high confidence labels with those of the unlabeled image to obtain the synthesized image.

[0052] Furthermore, taking the high-resolution scalp MRI dataset as an example, the input slices are randomly cropped to a size of 224×224, and then divided into labeled images. and unlabeled images The two groups have the same number.

[0053] right Applying weak enhancement and strong enhancement (strong enhancement, based on the weak enhancement's use of random cropping and random rotation, uses random grayscale changes) yields the following results: and ,in This represents an unlabeled image that has undergone weak enhancement. This indicates an unlabeled image that has undergone strong enhancement.

[0054] Labeled images Inputting the ADDA module yields a synthesized labeled image. .

[0055] Mark the amplitude of the image The data is fed into a deep reinforcement learning model, where each pixel is treated as an agent.

[0056] Calculate the side length W of the exchange region on the amplitude maps of the labeled and unlabeled images, and then... The central W×W region is assigned Corresponding area, and then The images are then fused and style-transferred to obtain the synthesized image through inverse Fourier transform. .in, This indicates that the process has been transformed. Superscript This indicates that it belongs to a labeled image, indicated by the superscript. This indicates that it belongs to an unlabeled image.

[0057] Will The data is fed into a segmentation network to obtain the segmentation results. The amplitude of the labeled image will be Feed the policy into the policy network P, and perform a fully connected operation on it using 1x1 convolutions to generate the policy.

[0058] Strategy Generate an action , It is the action of transferring the style of an unlabeled image to a labeled image. (for action space) It refers to the state, i.e., the predicted outcome. ( In a deep reinforcement learning model, the value network V (in the state space) provides the agent with the expected reward v for the current action based on the state feedback after the current action. The actual reward value is r. The agent adjusts its actions according to the advantage function, iterates continuously, and dynamically learns a scaling factor. This is used to control the size of the swap area, and is represented as:

[0059]

[0060] in, Indicates at time The scale factor. Indicates the first The image amplitude of each pixel is then calculated. The side length of the swapped region on the amplitude maps of the labeled and unlabeled images is then calculated using the following formula:

[0061]

[0062] in, The height of the input image.

[0063] Will The central W×W region is assigned Corresponding area, and then The images are then fused and style-transferred to obtain the synthesized image through inverse Fourier transform. The formula is expressed as:

[0064]

[0065] Will The data is fed into a segmentation network to obtain the segmentation results. For the predicted segmentation results and real labels Calculate cross-entropy loss As a reward, update the value network V:

[0066]

[0067] in, Indicates the first The real label of each pixel Indicates the first The predicted segmentation result for each pixel.

[0068] It should be noted that the task of the value network V is to analyze the current prediction results of the model. To evaluate the expected reward value that the agent's actions can obtain; to evaluate the predicted segmentation results and real labels Calculate the cross-entropy loss and use it as the true reward r to update the value network V.

[0069] S3: Perturb the input unlabeled image to obtain weakly perturbated and strongly perturbated images. Combine the synthesized image and the original image and input them into the encoder to extract image features step by step. Combine the features of the labeled image and the unlabeled image and input them into the decoder and classifier to train a semi-supervised image segmentation network.

[0070] Furthermore, the construction of the value network loss Strategy network loss ;Calculate the loss of the synthesized image using cross-entropy , will lose The segmentation network is used as a reward value to evaluate the style transfer enhancement of the image. The ability to segment and calculate value loss. The total loss of the reinforcement learning model is calculated.

[0071] Training a semi-supervised image segmentation network involves processing unlabeled images with weak perturbations. and strongly disturbed images The encoder is used to obtain features from weakly and strongly perturbated images, and low-order features from the weakly and strongly perturbated images are extracted respectively. and Then and Input phase alignment module.

[0072] Building a value network loss This is used for updating the value network. The loss of the synthesized image is calculated using cross-entropy. This value is used as a reward to evaluate the segmentation network's effect on style transfer enhancement of images. The ability to divide, thereby calculating the value loss. :

[0073]

[0074] Constructing a strategy network loss Used for updating the policy network, represented as:

[0075]

[0076] in, It is the dominant function. exp is Operation. The advantage function represents the range of advantages of the current action. If it is greater than 0, it means that the actual reward value is greater than the expected reward value, which is worth encouraging; otherwise, it should be suppressed.

[0077] The total loss of the reinforcement learning model is:

[0078]

[0079] in , , These are hyperparameters set based on experience.

[0080] It should be noted that an amplitude-aware deep reinforcement learning data augmentation strategy is proposed, which integrates Fourier transform into the deep reinforcement learning framework and adaptively performs style transfer enhancement on labeled images with confidence labels and different unlabeled images, thereby enabling the segmentation model to obtain better generalization ability through labeled images with diverse styles.

[0081] S4: Input the weakly perturbated image, the strongly perturbated image, and the synthetic image of the unlabeled image into the encoder to obtain low-order features, and process the low-order features of the unlabeled image through a phase alignment strategy.

[0082] Furthermore, the phase alignment strategy includes obtaining the phase alignment through Fourier transform. The phase spectrum is , The phase spectrum is Original image The phase spectrum is .

[0083] Will , and All inputs to the encoder (using ResNet101) yield four corresponding feature sets: and (Low-order and high-order features of labeled images). and (Synthesizing low-order and high-order features of labeled images) and (Low-order and high-order features of weakly enhanced unlabeled images) and (Strongly enhance low-order and high-order features of unlabeled images).

[0084] Will and Input the PHA module to calculate the phase consistency loss.

[0085] Low-order features of weak and strong perturbation images and The filtered features are obtained by feeding them into the multi-scale INC module, and the phase spectrum is obtained by applying Fourier transform. , Original image The phase spectrum is Calculate the phase consistency loss between images with weak and strong perturbations. :

[0086]

[0087] Here, i represents the value of the i-th pixel, which is constrained by the mean squared error loss.

[0088] Calculate the phase consistency loss between the weakly and strongly perturbed images and the original image. Align the perturbation image with the original image.

[0089]

[0090] The total phase alignment loss is expressed as:

[0091]

[0092] in , These are hyperparameters set based on experience.

[0093] It should be noted that a phase alignment strategy is introduced to guide the model to focus on structural components in the image, thereby effectively mitigating the interference of noise and artifacts. This strategy maintains structural consistency among the original image, the weakly enhanced image, and the strongly enhanced image by utilizing the phase information of the unlabeled image (because the phase components of the features contain structural information of the image).

[0094] S5: Input the high-level features of the perturbed unlabeled image into the cross-power spectrum correlation constraint module to obtain constraints on the segmentation results; add weights to calculate the total loss, perform backpropagation to update network parameters, and optimize the image segmentation network.

[0095] Furthermore, the cross-power spectrum correlation constraint module includes processing the low-level features of weakly and strongly enhanced unlabeled images input into the CPSC module.

[0096] The perturbed unlabeled image Advanced features The input is a Cross-Power Spectrum Correlation Constraint (CPSC) module. This CPSC module uses two classifiers with different initializations to predict the segmentation results of the input features, resulting in two logical outputs. The prediction results are subjected to Fourier transform. The normalized cross power spectrum is calculated for the frequency domain representation of the two different prediction results. The real part of the spectrum is then taken to generate a spatial domain correlation graph to capture global constraints. The spatial offset of the two prediction results is calculated and the offset is gradually corrected to align the two prediction results, thereby obtaining the correlation constraints on the segmentation results. The correlation constraints on the segmentation results are applied to the output of the original decoder to obtain the constrained prediction results and construct consistency constraints.

[0097] Will and Input the CPSC module to perform constraint calculations.

[0098] Will (Because both high-level features of weakly and strongly enhanced unlabeled images need to be processed by the CPSC module, we will only refer to unlabeled images here for general purposes.) The images are fed into two different classifiers, resulting in two logical outputs. .in, This represents the logical output obtained from the high-order features of the unlabeled image in classifier 1. This represents the logical output obtained from the high-order features of the unlabeled image in classifier 2.

[0099] For logical output and Using Fourier transform, we obtain their frequency domain representations. and .

[0100] calculate and Normalized cross power spectrum :

[0101]

[0102] in, express The associated conjugate matrix, for Prevent division by zero.

[0103] The backpropagation update of network parameters includes, for Perform an inverse Fourier transform to convert to the spatial domain, take the real part, and generate a spatial domain correlation graph. .

[0104] Find the two peaks in the correlation graph and calculate the spatial offset. .

[0105] Using spatial offset Align to After alignment Constructing a related graph , representing the probabilistic mean of the two segmentation results.

[0106] Using related graphs Adjust the decoder's logic output to obtain the constrained logic output; apply a softmax operation to the constrained logic output to obtain the prediction result. This makes the decoder's prediction Consistency alignment is performed between the constrained predictions, and the relevant loss is calculated; KL divergence is used for the unlabeled images with the two perturbations to reduce the gap between the predictions.

[0107] Simultaneously with the feature processing of the CPSC module, the corresponding high-order and low-order features are fused and fed into the decoder. Multi-scale convolution is used to reduce the dimensionality of the high-order features to be consistent with the low-order features. The two are then concatenated using a cat operation. After concatenation, the image is passed through a classifier to obtain a prediction. Bilinear interpolation is used to enlarge the image to the same size as the original image. The prediction result is obtained, and the total loss is calculated.

[0108] right Perform an inverse Fourier transform to convert to the spatial domain, take the real part, and generate a spatial domain correlation graph. :

[0109]

[0110] Find the two peaks in the correlation graph and calculate the spatial offset. :

[0111]

[0112] Using spatial offset Align to After alignment :

[0113]

[0114] Here, S represents a cyclic displacement operation.

[0115] Constructing a related graph , representing the probabilistic mean of the two segmentation results:

[0116]

[0117] in It is a softmax function applied along the category dimension.

[0118] Using related graphs Adjust the logic output of the decoder to obtain the constrained logic output:

[0119]

[0120] Correlation loss calculation: Apply a softmax operation to the constrained logical output to obtain the unlabeled image correlation prediction results. This makes the decoder's prediction Align the results with the constrained predictions and calculate the relevant losses:

[0121]

[0122] in, Indicates the first Relevance prediction results for unlabeled images of 1 pixel each. Indicates the first The decoder predicts each pixel.

[0123] In addition, KL divergence is used for the unlabeled images with the two perturbations to reduce the gap between predictions:

[0124]

[0125] in, This represents the weakly enhanced unlabeled image. The prediction result for each pixel. This represents the first unlabeled image after strong enhancement. The prediction results for each pixel are calculated. (This step is performed simultaneously with the cpsc module.) The corresponding high-order and low-order features are fused and fed into the decoder (using multi-scale convolution to reduce the dimensionality of the high-order features to be consistent with the low-order features, and then concatenating them using a cat operation. After concatenation, the images are passed through a classifier to obtain predictions, which are then enlarged to the same size as the original image using bilinear interpolation), resulting in four predictions: , , , . This represents the classifier prediction results for labeled images. This represents the classifier prediction result for the synthesized image. This represents the classifier prediction results for a weakly perturbated unlabeled image. This represents the classifier prediction results for a strongly perturbed unlabeled image. Finally, the total loss is calculated.

[0126] Finally, calculate the total loss:

[0127]

[0128] in These hyperparameters are set based on experience.

[0129] It should be noted that a cross-power spectral correlation module is proposed to capture global constraints and alleviate the misclassification problem of layered image structures. This module efficiently captures the spatial offset between the outputs of two different classifiers through cross-power spectral analysis, achieving this by finding its peak. The subsequently constructed correlation map is used to modulate high-level features to align the outputs of the two classifiers.

[0130] Example 2, Figures 2-4 As an embodiment of the present invention, a semi-supervised segmentation method based on amplitude-aware deep reinforcement learning is provided. In order to verify the beneficial effects of the present invention, scientific demonstration is carried out through economic benefit calculation and simulation experiments.

[0131] First, the dataset used in the invention comes from two sources:

[0132] The high-resolution MRI dataset of the scalp was obtained from a clinical hospital. It includes HR-MR scan images of the scalp from 54 patients. Each case includes two scans: one of the top of the head and one of the occipital region. Each scan image contains 18 slices, each 384×384 pixels. Six layers were manually labeled in this dataset: stratum corneum and granular layer (SCG), stratum spinosum and stratum basale (SSB), stratum papillary (PL), stratum reticularis (RL), subcutaneous fat layer with hair follicles (SFL-HF), and subcutaneous fat layer without hair follicles (SFL-NHF). After removing invalid slices, a total of 1240 slices were obtained. For training, 833 slices from 37 patients were selected, including 12 patients (245 slices) with 30% labeled data, 4 patients (83 slices) with 10% labeled data, and 2 patients (41 slices) with 5% labeled data. 4 patients (100 slices) were used for validation, and the remaining 11 patients (307 slices) were used for testing. The challenges of this dataset include significant noise, interlayer artifacts, and blurred layer boundaries.

[0133] The publicly available retinal OCT dataset is the MS dataset. It contains retinal OCT images from 35 subjects, including 14 healthy controls (HC) and 21 patients with multiple sclerosis (MS). Each retinal OCT contains 49 slices, with an image size of 496×1024 pixels, labeled with eight retinal layers: neural fiber layer (RNFL), ganglion cell layer and inner retinal layer (GCL+IPL), inner nuclear layer (INL), outer retinal layer (OPL), outer nuclear layer (ONL), inner photoreceptor segment (IS), outer photoreceptor segment (OS), and retinal pigment epithelium (RPE). Each slice is divided into two parts centered on the macula, resulting in 3430 retinal OCT images of size 384×512, randomly divided into 2940 training images, 196 validation images, and 294 test images. During training, five labeled images from different subjects were randomly used.

[0134] Experimental environment:

[0135] This invention trains and tests the model on an NVIDIA RTX 4090. The entire network is trained 200 times on the scalp HR-MR dataset and 100 times on the retina OCT dataset, with a learning rate of 0.0001. The Adam optimizer is used to train the segmentation model. The batch size is 6, and the batch size decays every 25 epochs with a factor of 0.9. DeepLabV3+ based on ResNet101 is used as the encoder.

[0136] The experimental results of this invention are as follows:

[0137] To quantitatively evaluate the performance of the method proposed in this invention, the Dice similarity coefficient (DSC) was selected as an evaluation metric to assess the segmentation performance of the network. Its calculation formula is as follows:

[0138]

[0139] in It represents the number of true positive segmentation pixels. It represents the number of segmented pixels that produce false positives. It represents the number of pixels that segmented the data as false negatives.

[0140] The present invention compares the segmentation results with existing semi-supervised models, as shown in Tables 1 and 2. Table 1 compares the present invention with other methods on the scalp HR-MR dataset, and Table 2 compares the present invention with other methods on the retinal dataset.

[0141] Table 1 Comparison of HR-MR dataset results for scalp

[0142]

[0143] As shown in Table 1, the present invention achieved the best performance in terms of average Dice score. Regarding Dice score, in the SCG, SSB, SFL-HF, and SFL-NHF layers of the scalp MRI dataset, the present invention outperformed the second-best method by 0.01, 2.02, 1.49, and 1.47, respectively, with 5% annotation. With 10% annotation, the present invention outperformed the second-best method by 1.32, 3.89, 0.92, 2.26, and 1.42, respectively, in Dice score for each layer. With 30% annotation, the present invention outperformed the second-best method by 0.25, 1.73, 1.54, 0.85, and 1.79, respectively, in Dice score for the first 5 layers.

[0144] Table 2 Comparison of Results for Retina Dataset

[0145]

[0146] Table 2 shows the performance of this invention and the state-of-the-art (SOTA) method on the retinal OCT dataset for retinal layer segmentation. This invention achieves the best results across seven layers, with an average Dice coefficient 1.87 higher than the second-best method. Specifically, the Dice coefficients for RNFL, GCIP, INL, ONL, IS, OS, and RPE layers are 1.93, 3.43, 3.02, 0.31, 1.81, 0.56, and 1.14 higher than the second-best method, respectively. Clearly, most methods achieve high accuracy when segmenting thinner layers such as IS and OS; in contrast, this invention not only performs well in these layers but also outperforms the fully supervised method by 0.5 on the OS layer.

[0147] Figure 3 and Figure 4 These are segmentation results comparing the scalp HR-MR dataset and the retinal dataset with other methods.

[0148] like Figure 3 As shown, a visual comparison of the various methods was performed using only 10% of the labeled image. The methods performed poorly when segmenting thinner layers, resulting in blurred boundaries and low contrast between adjacent layers. The Crcfp method led to oversegmentation of layers, while BCP, ABD (BCP), and SDCL tended to produce undersegmentation in layers with low contrast. UniMatch lost several thin layers, and CorrMatch also failed to accurately restore the boundary details of the thinnest layer. In contrast, this invention not only maintains the structure of layer boundaries but also preserves the detailed information of thin layers.

[0149] Figure 4 This paper presents a visual comparison between the present invention and related methods. Layer segmentation in the macular region presents certain challenges. Crcfp exhibits oversegmentation and undersegmentation at multiple layers, while BCP, SDCL, and ABD (BCP) also suffer from undersegmentation at multiple layers. Although UniMatch and CorrMatch produce relatively complete segmentation results overall, they perform poorly in thinner layers (such as RNFL), resulting in the loss of detail. In contrast, the present invention provides clear boundaries across each layer structure, effectively mitigating the oversegmentation and undersegmentation problems across multiple layers.

[0150] Example 3, an embodiment of the present invention, provides a semi-supervised segmentation system based on amplitude-aware deep reinforcement learning, including an ADDA module, a phase alignment PHA module, and a cross-power spectrum correlation CPSC module.

[0151] The ADDA module is used to combine Fast Fourier Transform and Inverse Fourier Transform to achieve adaptive image style transfer.

[0152] The phase alignment (PHA) module is used to enable the model to focus only on the structural information of the image at the feature level through a phase consistency strategy.

[0153] The Cross Power Spectrum Correlation (CPSC) module is used to capture global constraints through cross power spectrum calculation, thereby optimizing prediction alignment and consistency in layer segmentation tasks.

[0154] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this invention, or the part that contributes to the prior art, or a part of the technical solution, can be embodied in the form of a software product. This computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods of the various embodiments of this invention. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0155] The logic and / or steps represented in the flowchart or otherwise described herein, for example, can be considered as a sequenced list of executable instructions for implementing logical functions, and can be embodied in any computer-readable medium for use by, or in conjunction with, an instruction execution system, apparatus, or device (such as a computer-based system, a processor-including system, or other system that can fetch and execute instructions from, an instruction execution system, apparatus, or device). For the purposes of this specification, "computer-readable medium" can be any means that can contain, store, communicate, propagate, or transmit programs for use by, or in conjunction with, an instruction execution system, apparatus, or device.

[0156] More specific examples (a non-exhaustive list) of computer-readable media include: electrical connections (electronic devices) having one or more wires, portable computer disk drives (magnetic devices), random access memory (RAM), read-only memory (ROM), erasable and editable read-only memory (EPROM or flash memory), fiber optic devices, and portable optical disc read-only memory (CDROM). Furthermore, computer-readable media can even be paper or other suitable media on which programs can be printed, because programs can be obtained electronically, for example, by optically scanning the paper or other media, followed by editing, interpreting, or otherwise processing as necessary, and then stored in computer memory.

[0157] It should be understood that various parts of the present invention can be implemented using hardware, software, firmware, or a combination thereof. In the above embodiments, multiple steps or methods can be implemented using software or firmware stored in memory and executed by a suitable instruction execution system. For example, if implemented in hardware, as in another embodiment, it can be implemented using a combination of any of the following techniques known in the art: discrete logic circuits having logic gates for implementing logical functions on data signals, application-specific integrated circuits (ASICs) having suitable combinational logic gates, programmable gate arrays (PGAs), field-programmable gate arrays (FPGAs), etc. It should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and are not intended to limit it. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can be made to the technical solutions of the present invention without departing from the spirit and scope of the technical solutions of the present invention, and all such modifications or substitutions should be covered within the scope of the claims of the present invention.

Claims

1. A semi-supervised segmentation method based on amplitude-aware deep reinforcement learning, characterized in that, include: We acquired scalp HR-MR images as a dataset and labeled the image segmentation targets. The input labeled and unlabeled images are decomposed to obtain style and structural information. Through a deep reinforcement learning strategy, the style features of the labeled image with high confidence labels and the unlabeled image are adaptively swapped to obtain the synthetic image. The input unlabeled image is perturbed to obtain weakly perturbed and strongly perturbed images. The synthesized image and the original image are then input into the encoder to extract image features step by step. The features of labeled and unlabeled images are jointly input into the decoder and classifier to train a semi-supervised image segmentation network. Weakly perturbed images, strongly perturbed images, and synthetic images of unlabeled images are input into the encoder to obtain low-order features, and the low-order features of the unlabeled images are processed through a phase alignment strategy. The high-level features of the perturbed unlabeled image are input into the cross-power spectrum correlation constraint module to obtain constraints on the segmentation results; weights are added to calculate the total loss, and the network parameters are updated through backpropagation to optimize the image segmentation network.

2. The semi-supervised segmentation method based on amplitude-aware deep reinforcement learning as described in claim 1, characterized in that: The annotation of image segmentation targets includes having professionals annotate the segmentation targets in the data and dividing it into training and testing sets; The image decomposition includes processing the labeled image. and unlabeled images The labeled images were obtained by performing frequency domain decomposition using Fourier transform. The amplitude is Phase is Unlabeled images The amplitude is .

3. The semi-supervised segmentation method based on amplitude-aware deep reinforcement learning as described in claim 2, characterized in that: The deep reinforcement learning strategy includes a policy network P in the deep reinforcement learning model based on the policy. Generate an action , This represents the action of transferring the style of an unlabeled image to a labeled image, where... For the action space, For state space; The state represents the input received by the model at time t. ; Mark the amplitude of the image The data is fed into a deep reinforcement learning model, where each pixel is treated as an agent. The value network V in the deep reinforcement learning model evaluates the state after the current action, resulting in a predicted reward value v, while the actual reward value r is obtained from environmental feedback. The agent iterates by adjusting its actions according to the advantage function, dynamically learning a scaling factor. ,use Control the size of the switching area; Calculate the side length W of the exchange region on the amplitude maps of the labeled and unlabeled images, and then... The central W×W region is assigned Corresponding area, and then The images are then fused and style-transferred to obtain the synthesized image through inverse Fourier transform. ; Will The data is fed into a segmentation network to obtain the segmentation results. The task of the value network V is to estimate the expected reward value that the agent can obtain from its actions, which is the model's current prediction. ; For the predicted segmentation results and real labels Calculate cross-entropy loss The real reward r is used to update the value network V.

4. The semi-supervised segmentation method based on amplitude-aware deep reinforcement learning as described in claim 3, characterized in that: The semi-supervised training of the image segmentation network includes constructing a value network loss. Strategy network loss ;Calculate the loss of the synthesized image using cross-entropy , will lose The segmentation network is used as a reward value to evaluate the style transfer enhancement of the image. The ability to segment and calculate value loss. ;Calculate the total loss of the reinforcement learning model; Weakly perturbed images of unlabeled images and strongly perturbated images The encoder is used to obtain features from weakly perturbed and strongly perturbed images, and low-order features from the weakly perturbed and strongly perturbed images are extracted respectively. and Then and Input phase alignment module.

5. The semi-supervised segmentation method based on amplitude-aware deep reinforcement learning as described in claim 4, characterized in that: The phase alignment strategy includes obtaining it through Fourier transform. The phase spectrum is , The phase spectrum is Original image The phase spectrum is ; Low-order features of weak and strong perturbation images and The filtered features are fed into the multi-scale INC module to obtain the filtered features, and the phase spectrum is obtained by applying Fourier transform. , Calculate the phase consistency loss between images with weak and strong perturbations. Align images with weak and strong perturbations.

6. The semi-supervised segmentation method based on amplitude-aware deep reinforcement learning as described in claim 5, characterized in that: The cross-power spectrum correlation constraint module includes processing the low-level features of weakly and strongly enhanced unlabeled images into the CPSC module. The perturbed unlabeled image Advanced features The input is a Cross-Power Spectrum Correlation Constraint (CPSC) module. This CPSC module uses two classifiers with different initializations to predict the segmentation results of the input features, resulting in two logical outputs. and The prediction results are subjected to Fourier transform. The normalized cross power spectrum is calculated for the frequency domain representation of the two different prediction results. The real part of the spectrum is then taken to generate a spatial domain correlation graph to capture global constraints. The spatial offset of the two prediction results is calculated and the offset is gradually corrected to align the two prediction results, thereby obtaining the correlation constraints on the segmentation results. The correlation constraints on the segmentation results are applied to the output of the original decoder to obtain the constrained prediction results and construct consistency constraints.

7. The semi-supervised segmentation method based on amplitude-aware deep reinforcement learning as described in claim 6, characterized in that: The process of updating network parameters via backpropagation includes, for Perform an inverse Fourier transform to convert to the spatial domain, take the real part, and generate a spatial domain correlation graph. : Find the two peaks in the correlation graph and calculate the spatial offset. ; Using spatial offset Align to After alignment Constructing a related graph , representing the probabilistic mean of the two segmentation results; Using related graphs Adjust the decoder's logic output to obtain the constrained logic output; apply a softmax operation to the constrained logic output to obtain the prediction result. This makes the decoder's prediction Consistency alignment is performed with the constrained predictions, and the relevant loss is calculated; KL divergence is used for the unlabeled images with the two perturbations to reduce the gap between the predictions. Simultaneously with the feature processing of the CPSC module, the corresponding high-order and low-order features are fused and fed into the decoder. Multi-scale convolution is used to reduce the dimensionality of the high-order features to be consistent with the low-order features. The two are then concatenated using a cat operation. After concatenation, the image is passed through a classifier to obtain the prediction. Bilinear interpolation is then used to enlarge the image to the same size as the original image. Obtain the prediction results and calculate the total loss.

8. A system employing a semi-supervised segmentation method based on amplitude-aware deep reinforcement learning as described in any one of claims 1 to 7, characterized in that: Includes ADDA module, phase alignment PHA module, and cross power spectrum correlation CPSC module; The ADDA module is used to combine fast Fourier transform and inverse Fourier transform to achieve adaptive image style transfer. The phase alignment (PHA) module is used to enable the model to focus only on the structural information of the image at the feature level through a phase consistency strategy. The Cross Power Spectrum Correlation (CPSC) module is used to capture global constraints through cross power spectrum calculation, thereby optimizing prediction alignment and consistency in layer segmentation tasks.

9. A computer device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the semi-supervised segmentation method based on amplitude-aware deep reinforcement learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the steps of the semi-supervised segmentation method based on amplitude-aware deep reinforcement learning as described in any one of claims 1 to 7.

Images