Method for subjective cognitive decline state detection based on generated image data

Abstract

The invention discloses a subjective cognitive decline state detection method based on generated image data, and the method comprises the steps: firstly obtaining two data sets of an image generation sample and a disease state detection sample based on a disclosed Alzheimer's disease data set, and carrying out the preprocessing; secondly, constructing an image generation deep learning model, using the preprocessed image generation sample data set as input data, training the model, and using the trained image generation model to complement a disease state detection sample data set; and finally, constructing a disease state detection deep learning model, inputting data in the complemented disease state detection sample data set, outputting and obtaining a detection result of the subjective cognitive decline state, and performing training. The PET image generated by the method can effectively supplement limitation of MRI data, so that better performance is shown in illness state detection, and a new research direction is provided for future multi-modal medical image analysis.

Description

Technical Field

[0001] The present invention belongs to the field of medical image processing and analysis, and in particular relates to a method for detecting a subjective cognitive decline state based on generated image data. Background Art

[0002] Alzheimer's disease (AD) is an irreversible neurodegenerative disease. Subjective cognitive decline (SCD) is considered to be the earliest clinical stage of AD. SCD patients may progress to mild cognitive impairment (MCI) or even AD within a few years. Therefore, it is particularly important to identify SCD patients in progression as early as possible and intervene in time.

[0003] The condition of SCD patients can be detected through a variety of neuroimaging techniques. Magnetic resonance imaging (MRI) and positron emission tomography (PET) are two key technologies that provide complementary structural and functional brain data and have become effective means of detecting brain abnormalities in SCD patients. However, obtaining comprehensive multimodal neuroimaging data (especially PET data) still faces great challenges, mainly due to the high cost of PET imaging and patients' concerns about radiation. Therefore, how to complete the missing imaging data at a low cost has become a key issue that needs to be solved in the diagnosis of SCD progression.

[0004] Current SCD progression diagnostic methods are limited by the availability of neuroimaging data and mainly rely on a single MRI modality, making it difficult to fully utilize the complementary information of multimodal neuroimaging. In addition, existing deep learning models often have difficulty accurately extracting complex features when processing three-dimensional neuroimaging, mainly due to the multidimensionality and complexity of neuroimaging data. Summary of the invention

[0005] In response to the above problems, the present invention proposes a method for detecting the state of subjective cognitive decline based on generated imaging data, which can generate high-quality PET images from existing MRI images and complete the missing imaging modalities, thereby providing more comprehensive auxiliary information for diagnosis and improving the accuracy of diagnosis. At the same time, in order to improve the feature extraction capability of the deep learning model, the present invention introduces a three-domain attention mechanism, which integrates the features of the image domain, channel domain, and frequency domain, improves the feature extraction effect of the model from multiple dimensions, and ultimately improves the overall performance of the model.

[0006] The method for predicting the progression of subjective cognitive decline based on generated imaging data comprises the following steps:

[0007] Step 1: Based on the public Alzheimer's disease dataset, obtain two datasets: image generation samples and disease state detection samples.

[0008] In the Alzheimer's Disease Neuroimaging Initiative (ADNI) public database, magnetic resonance imaging (MRI) and positron emission tomography (PET) image data of patients who met the criteria were screened and obtained, and the patients' disease status classification labels were determined based on clinical evaluation and subjective psychological test scores (such as the Mini-Mental Inventory).

[0009] Through the above steps, two datasets are obtained: image generation samples and disease state detection samples. The image generation sample dataset contains matching MRI and PET image pairs, while the disease state detection sample dataset contains the patient's disease state classification label and MRI image.

[0010] Step 2: Preprocess the image data in the two data sets of image generation samples and disease state detection samples.

[0011] The present invention proposes a series of standardized preprocessing processes for image data in image generation sample data sets and disease state detection sample data sets. First, all MRI and PET images are subjected to skull stripping using SynthStrip software. Subsequently, the MRI images are subjected to intensity inhomogeneity correction using FreeSurfer software. Next, each PET image is spatially matched with its corresponding MRI image by rigid registration. Then, all MRI and PET images are standardized to the Montreal Neurological Institute (MNI) space using Advanced Normalization Tools (ANTs) to ensure spatial consistency of each pair of MRI and PET images.

[0012] After completing the above preprocessing steps, all MRI and PET images were resampled to the MNI template size (182×218×182 voxels) and the isotropic resolution was set to 1 mm. These processed and normalized image data can ensure the comparability of image information between different patients, thereby providing standardized input data for subsequent deep learning models.

[0013] Step 3: Build an image generation deep learning model, use the preprocessed image generation sample dataset as input data, and train the image generation model.

[0014] The image generation model constructed in step 3 uses the 3D conditional generative adversarial network (3D conditionalGAN) as the basic model. The model consists of two main components: the generator and the discriminator. The generator uses the U-Net framework with jump connections and introduces a triple attention mechanism to enhance feature extraction capabilities and overall performance. Similarly, the discriminator based on the convolutional neural network (CNN) architecture also adds a triple attention mechanism, which improves the ability to distinguish real images from synthetic images and enhances the effect of adversarial training.

[0015] The generator adopts a U-shaped path structure of encoding layer-intermediate layer-decoding layer, which includes 1 encoding layer, 7 intermediate layers and 1 decoding layer. The encoding layer contains 1 leaky ReLU layer and 1 convolution layer to increase the number of input feature channels. Each intermediate layer consists of a downsampling module and an upsampling module. The downsampling module includes a leaky ReLU layer, a convolution layer and an instance normalization layer, and the upsampling module consists of a ReLU layer, a deconvolution layer and an instance normalization layer. The upsampling and downsampling modules of the same intermediate layer are connected by skip connections. The data is first downsampled through each layer and then upsampled back layer by layer. Finally, the 2nd to 5th intermediate layers integrate triple attention modules to further improve the performance of the model. The decoding layer starts with a leaky ReLU layer, followed by a deconvolution layer to restore the number of feature channels to the original size, and a tanh activation layer is used to output the final result.

[0016] The discriminator consists of 5 downsampling layers, each of which contains 1 convolutional layer, 1 instance normalization layer, and 1 leaky ReLU activation layer; the 4th layer adds a triple attention module to enhance the model performance. The 5th layer adds an additional convolutional layer to restore the number of feature channels to the original size and generate the final discriminant result.

[0017] In the model design, a variety of loss functions are used to optimize the training image generation model, and the quality of the generated image is improved by capturing complex relationships and ensuring the consistency between the source modality and the target modality, including GAN loss, feature matching loss, brain region RMSE loss, and focal frequency loss. GAN loss is the core component, and a conditional GAN ​​loss function is designed as the basic loss function; feature matching loss is used to stabilize the training process and reduce the inconsistency between the generator and the discriminator; brain region RMSE loss calculates the root mean square error (RMSE) between the masked generated image and the target image, concentrates the error calculation on the brain region, and excludes irrelevant background noise; focal frequency loss is to solve the missing problem in the frequency domain, expand to three-dimensional space, adaptively emphasize frequency components that are difficult to synthesize, and supplement the existing spatial loss by retaining important frequency information that is difficult for neural networks to capture. The final objective function can be expressed by the following formula:

[0018]

[0019] Among them, G and D represent the generator and discriminator respectively. cGAN (G,D) and λ adv is the conditional GAN ​​loss and its weight coefficient. L2(D) and μ D is the discriminator L2 regularization loss and its weight coefficient. L L1 (G) and λ L1 is the L1 loss of the generator and its weight coefficient. L FM (G,D) and λ FM is the feature matching loss and its weight coefficient. B-rmse (G) and λ B-rmse is the RMSE loss of the brain region and its weight coefficient. FF (G,D) and λ FF is the focal frequency loss and its loss function. L2(G) and μ G is the generator L2 regularization loss and its weight coefficient.

[0020] Use the previously preprocessed image generation sample dataset as input data and train the image generation model for several rounds until the model converges.

[0021] Step 4: Use the trained image generation model to complete the disease state detection sample data set.

[0022] Use the image generation model trained in the previous step to generate corresponding PET images for the MRI images in the disease state detection sample dataset, thereby completing the disease state detection sample dataset. The working process can be described by the following formula:

[0023]

[0024] in Represents the input MRI image x M Generated PET images. M and I P They represent the domain of MRI images and the domain of PET images respectively. It is an image generation model constructed to generate PET images from MRI images.

[0025] Step 5: Build a deep learning model for disease state detection, input the data in the completed disease state detection sample data set, output the detection results of subjective cognitive decline state, and conduct training.

[0026] The disease state detection model constructed in step 5 uses a three-dimensional convolutional neural network (3D CNN) as the basic model. The disease state prediction model consists of 5 downsampling layers, each of which contains a convolutional layer, a batch normalization layer, and a leaky ReLU activation layer. The fourth layer adds a triple attention module to enhance the model performance. The fifth layer adds an additional fully connected layer and a Sigmoid activation layer to generate and output the final classification label.

[0027] Due to the class imbalance in the classification task dataset, focal loss is used as the loss function in the disease state detection model. This method assigns different weights to different categories, thereby giving greater influence to categories with fewer samples during training. The loss function can be expressed as follows:

[0028]

[0029] Where N represents the total number of subjects, y i is the true classification label of the i-th subject, p i is the predicted probability of the i-th subject given by the model. is category y i is a weight factor used to solve the class imbalance problem. γ is a focusing parameter used to reduce the contribution of easy-to-classify samples to the loss and emphasize difficult-to-classify samples.

[0030] The data in the completed disease state detection sample data set is used as input data to train the disease state detection model until the disease state detection model converges.

[0031] Beneficial effects of the present invention

[0032] This paper proposes a novel comprehensive method framework, which combines an image synthesis network based on a three-dimensional generative adversarial network (3DGAN) and a classification network based on a three-dimensional convolutional neural network (3D CNN), aiming to improve the detection effect of the condition status of patients with subjective cognitive decline (SCD). The method first uses 3D conditional GAN ​​to generate high-quality PET image data to supplement the complementary information that may be missing in MRI data, thereby enhancing the overall expression ability of multimodal image data. Then, the classification network based on 3D CNN makes full use of the generated multimodal image information to further improve the accuracy of SCD disease progression prediction.

[0033] Experiments on the Alzheimer's Disease Neuroimaging Initiative (ADNI) dataset verified the effectiveness of the framework. Specific results show that this method achieved an excellent performance of an average structural similarity index (SSIM) value of 0.882 in the cross-modal image synthesis task; and in the SCD disease progression prediction task, the area under the receiver operating characteristic curve (AUC) value reached 0.7633, significantly better than other comparison methods. The experimental results show that the generated PET images can effectively supplement the limitations of MRI data, thereby showing better performance in disease status detection.

[0034] In general, this method can overcome the difficulty of multimodal medical image acquisition and improve the accuracy of disease state detection by generating missing medical image modalities. This framework provides a new research direction for future multimodal medical image analysis and has potential for broad application prospects. BRIEF DESCRIPTION OF THE DRAWINGS

[0035] Figure 1 Generate a deep learning model for the image constructed by the present invention;

[0036] Figure 2 A deep learning model for disease state detection constructed for the present invention. DETAILED DESCRIPTION

[0037] The method for detecting a subjective cognitive decline state based on generated image data comprises the following steps:

[0038] Step 1: Based on the public Alzheimer's disease dataset, obtain two datasets: image generation samples and disease state detection samples.

[0039] In the Alzheimer's Disease Neuroimaging Initiative (ADNI) public database, magnetic resonance imaging (MRI) and positron emission tomography (PET) image data of patients who met the criteria were screened and obtained, and the patients' disease status classification labels were determined based on clinical evaluation and subjective psychological test scores (such as the Mini-Mental Inventory).

[0040] Through data retrieval and screening, a total of 353 patients with subjective cognitive decline (SCD) were identified in the ADNI dataset, and 627 pairs of T1MRI and amyloid PET imaging data of these patients were collected. This study aims to predict the disease progression of SCD patients and define the disease progression status of patients using a 60-month time window. Specifically, if a SCD patient develops mild cognitive impairment (MCI) within 60 months, he or she is classified as progressive SCD (pSCD); on the contrary, if the patient does not undergo disease transformation during this period, he or she is considered to be stable SCD (sSCD). Due to the lack of follow-up data for some patients, only 27 patients were classified as pSCD and 75 patients were classified as sSCD according to the above criteria.

[0041] Through the above steps, two datasets are obtained: image generation samples and disease state detection samples. The image generation sample dataset contains matching MRI and PET image pairs, while the disease state detection sample dataset contains the patient's disease state classification label and MRI image.

[0042] Step 2: Preprocess the image data in the two data sets of image generation samples and disease state detection samples.

[0043] Step 3: Build a deep learning model for image generation, such as Figure 1 As shown, the preprocessed image generation sample dataset is used as input data to train the image generation model.

[0044] Step 4: Use the trained image generation model to complete the disease state detection sample data set.

[0045] Step 5: Build a deep learning model for disease state detection, such as Figure 2 As shown, the data in the completed disease state detection sample data set is input, the detection result of the subjective cognitive decline state is output, and training is performed.

[0046] Example:

[0047] (1) Image generation model performance evaluation

[0048] In order to evaluate the performance of the proposed image generation network, the average structural similarity index (SSIM) and the average peak signal-to-noise ratio (PSNR) are used as evaluation indicators. SSIM measures the consistency of visual perception by analyzing the similarity of image brightness, contrast and structure, while PSNR evaluates the accuracy of image reconstruction by quantifying pixel-level errors. These two indicators can comprehensively evaluate the quality of generated images from different angles.

[0049] The proposed model is compared with three other mainstream models (3D pix2pix, SC-GAN, Res-GAN, TAC-GAN), and the experimental results are summarized in Table 1:

[0050] Table 1

[0051]

[0052]

[0053] According to the experimental results in the table, the image generation deep learning model proposed in the present invention achieved the highest SSIM score (0.882) among all models, significantly better than Res-GAN (0.879), 3D pix2pix (0.864) and SC-GAN (0.854), indicating that it has better performance in maintaining the structural integrity and detail features of the original image. At the same time, the PSNR value of TAC-GAN also reached 26.508, which also exceeded Res-GAN (26.326), 3D pix2pix (26.224) and SC-GAN (25.877), showing its outstanding performance in suppressing noise. These results show that TAC-GAN achieves the best balance between feature retention and noise suppression, and exhibits superior performance in generating high-quality images.

[0054] In general, the image generation deep learning model proposed in this paper outperforms other models in terms of structural similarity and image reconstruction accuracy, verifying its effectiveness and superiority in image generation tasks.

[0055] (2) Performance evaluation of disease state detection models

[0056] In order to comprehensively evaluate the performance of the disease state detection model, five evaluation indicators were used: Balanced Accuracy (BAC), Sensitivity (SEN), Specificity (SPE), Area Under the Curve (AUC) and F1 Score (F1S). These indicators measure the classification effect of the model from different angles. Generally speaking, the higher the value of each indicator, the better the performance of the model.

[0057] The proposed model is compared in detail with the methods proposed by other studies, and the experimental results are shown in Table 2 below:

[0058] Table 2

[0059]

[0060] From the results, the deep learning model for disease state detection proposed in the present invention is significantly superior to other methods in three core indicators: area under the receiver operating characteristic curve AUC (0.7633), balanced accuracy BAC (0.7923) and sensitivity SEN (0.8889). Among them, the AUC value is 0.7633, showing the excellent ability of the model in distinguishing the disease progression status of different patients (pSCD and sSCD); and the sensitivity value reaches 0.8889, indicating that the model has a higher detection rate when identifying pSCD patients (i.e. SCD patients whose disease will progress). In addition, although it did not reach the highest values ​​in specificity (SPE) and F1 score (F1S) (0.6957 and 0.6667, respectively), compared with other advanced methods, its results are still close to the optimal level (SPE is up to 0.7308 and F1S is up to 0.6770).

[0061] In summary, the disease status detection deep learning model proposed in the present invention outperforms other mainstream methods in overall performance, especially in terms of AUC and sensitivity, showing strong classification ability and stability, proving its effectiveness and advantages in detecting the disease status of SCD patients.

Claims

1. A method for detecting a subjective cognitive decline state based on generated image data, characterized in that: The following steps are involved: Step 1: Based on the public Alzheimer's disease dataset, obtain two datasets: image generation samples and disease state detection samples; Step 2: preprocessing the image data in the two data sets of image generation samples and disease state detection samples; Step 3: Build an image generation deep learning model, use the preprocessed image generation sample dataset as input data, and train the image generation model; Step 4: Use the trained image generation model to complete the disease state detection sample data set; Step 5: Build a deep learning model for disease state detection, input the data in the completed disease state detection sample data set, output the detection results of subjective cognitive decline state, and conduct training.

2. The method for detecting a subjective cognitive decline state based on generated image data according to claim 1, characterized in that: The specific implementation process of step 1 is as follows: Obtain MRI image data and positron emission tomography (PET) image data of patients with Alzheimer's disease, and determine the patient's disease status classification label based on clinical evaluation and subjective psychological test scores; Two data sets are obtained: image generation samples and disease state detection samples. The image generation sample data set contains matching MRI and PET image pairs, while the disease state detection sample data set contains the patient's disease state classification labels and MRI images.

3. The method for detecting a subjective cognitive decline state based on generated image data according to claim 2, characterized in that: The specific process of pretreatment described in step 2 is as follows: First, all MRI and PET images were processed with skull stripping; Secondly, the MRI images were corrected for intensity inhomogeneity; Then, each PET image is spatially matched to its corresponding MRI image through rigid registration; Finally, all MRI and PET images were normalized to the Montreal Neurological Institute MNI space using the advanced registration tool ANTs; After completing the above preprocessing steps, all MRI and PET images were resampled to the MNI template size and the isotropic resolution was set to 1 mm.

4. The method for detecting a subjective cognitive decline state based on generated image data according to claim 1, characterized in that: The specific implementation process of the image generation model is as follows: A three-dimensional conditional generative adversarial network is used as the basic model, which consists of two components: a generator and a discriminator. The generator uses a U-Net framework with skip connections and introduces a triple attention mechanism. The discriminator based on the convolutional neural network (CNN) architecture also incorporates a triple attention mechanism. A variety of loss functions are used to optimize the training image generation model, including GAN loss, feature matching loss, brain region RMSE loss, and focal frequency loss; the final objective function is expressed as follows: Among them, G and D represent the generator and discriminator respectively; L cGAN (G,D) and λ adv is the conditional GAN ​​loss and its weight coefficient; L2(D) and μ D is the discriminator L2 regularization loss and its weight coefficient; L L1 (G) and λ L1 is the L1 loss of the generator and its weight coefficient; L FM (G,D) and λ FM is the feature matching loss and its weight coefficient; L B-rmse (G) and λ B-rmse is the RMSE loss of the brain region and its weight coefficient; L FF (G,D) and λ FF is the focal frequency loss and its loss function; L2(G) and μ G is the generator L2 regularization loss and its weight coefficient; Use the previously preprocessed image generation sample dataset as input data and train the image generation model for several rounds until convergence.

5. The method for detecting a subjective cognitive decline state based on generated image data according to claim 4, characterized in that: The generator adopts a U-shaped path structure of encoding layer-intermediate layer-decoding layer, including 1 encoding layer, 7 intermediate layers and 1 decoding layer; The encoding layer contains 1 leaky ReLU layer and 1 convolution layer; each intermediate layer consists of a downsampling module and an upsampling module. The downsampling module includes a leaky ReLU layer, a convolution layer, and an instance normalization layer, and the upsampling module consists of a ReLU layer, a deconvolution layer, and an instance normalization layer. The upsampling and downsampling modules of the same intermediate layer are connected through jump connections; the data is first downsampled through each layer, and then upsampled back layer by layer; the last 2 to 5 intermediate layers integrate triple attention modules; the decoding layer starts with a leaky ReLU layer, and then the number of feature channels is restored to the original size through a deconvolution layer, and the final result is output using a tanh activation layer; The discriminator consists of 5 downsampling layers, each of which contains 1 convolutional layer, 1 instance normalization layer and 1 leaky ReLU activation layer; the 4th layer adds a triple attention module; the 5th layer adds an additional convolutional layer to restore the number of feature channels to the original size and generate the final discrimination result.

6. The method for detecting a subjective cognitive decline state based on generated image data according to claim 5, characterized in that: The specific implementation process of step 4 is as follows: Use the trained image generation model to generate corresponding PET images for the MRI images in the disease state detection sample dataset to complete the disease state detection sample dataset, which can be expressed as follows: in Represents the input MRI image x M Generated PET images; I M and I P They represent the domain of MRI images and the domain of PET images, respectively; It is an image generation model constructed to generate PET images from MRI images.

7. The method for detecting a subjective cognitive decline state based on generated image data according to claim 6, characterized in that: The specific implementation process of step 5 is as follows: The disease state detection model uses a three-dimensional convolutional neural network as the basic model; the disease state prediction model consists of 5 downsampling layers, each of which contains a convolutional layer, a batch normalization layer, and a leaky ReLU activation layer; the fourth layer adds a triple attention module, and the fifth layer adds an additional fully connected layer and a Sigmoid activation layer to generate and output the final classification label; In the disease state detection model, focal loss is used as the loss function; the loss function is expressed by the following formula: Where N represents the total number of subjects, y i is the true classification label of the i-th subject, p i is the predicted probability of the i-th subject given by the model; is category y i The weight factor of γ is the focusing parameter; The data in the completed disease state detection sample data set is used as input data to train the disease state detection model until the disease state detection model converges.

Images