Artificial Intelligence-Based Nuclear Medicine Image Prediction System

Abstract

This invention relates to the field of nuclear medicine imaging technology and discloses an artificial intelligence-based nuclear medicine image prediction system. The system includes a dynamic image acquisition layer, a deep feature extraction layer, a cross-modal temporal fusion layer, an intelligent prediction decision layer, and a dynamic feedback optimization layer. The system acquires image data through multimodal temporal sampling, generates a dynamic functional metabolic atlas through adaptive noise reduction, extracts deep features using 3D convolution and dilated convolution to generate a multi-scale semantic feature atlas, establishes a spatiotemporal correlation between static and dynamic data, and fuses features using an attention mechanism and gated recurrent units to generate lesion risk probabilities. Based on reinforcement learning, the risk probabilities are transformed into clinical prediction results, which are output through a standardized interface. The system tracks application status in real time, calculates deviations, and dynamically optimizes the model until the prediction efficacy index stabilizes. This system improves the comprehensiveness and accuracy of nuclear medicine image prediction, enhances its alignment with clinical needs, and achieves continuous model optimization.

Description

Technical Field

[0001] This invention relates to the field of nuclear medicine imaging technology, specifically to a nuclear medicine image prediction system based on artificial intelligence. Background Technology

[0002] Nuclear medicine imaging technology, as an important tool in clinical diagnosis, can reflect the functional state and physiological processes of tissues and organs by tracking the distribution and metabolism of radioactive tracers in the body. It plays a crucial role in the early diagnosis and efficacy evaluation of various diseases, including tumors, cardiovascular diseases, and neurological disorders. However, with the continuous development of medical imaging technology, the complexity and information content of nuclear medicine imaging data are constantly increasing, and traditional image analysis methods are gradually revealing many limitations.

[0003] The interpretation of nuclear medicine images relies heavily on the physician's experience and judgment, which is highly subjective and inconsistent. Different physicians may analyze the same image differently, especially in the early stages of disease when imaging features are not obvious, easily leading to missed diagnoses or misdiagnoses. Furthermore, nuclear medicine images typically contain dynamic temporal information, requiring comprehensive analysis in conjunction with static structural images. However, manual processing struggles to accurately capture the spatiotemporal correlation between dynamic function and static structure, limiting the accuracy of predicting disease progression.

[0004] While some existing automated analysis systems attempt to incorporate computer-aided diagnostic techniques, most are limited to static analysis of single-modal images, failing to effectively integrate dynamic functional metabolic data with static structural data. For example, some systems only extract features from static images, ignoring the dynamic process of tracer changes in vivo over time, making it difficult to reflect the dynamic development of lesions. Other systems, although involving dynamic data processing, lack spatiotemporal correlation modeling of multimodal data, resulting in insufficient effectiveness of feature fusion and affecting the reliability of prediction results.

[0005] Existing predictive models have weak adaptability and generalization capabilities, making it difficult to dynamically optimize them based on actual clinical applications. In real-world clinical scenarios, individual differences among patients, fluctuations in equipment parameters, and the complex manifestations of diseases can all lead to a decline in the predictive performance of the models. Traditional models typically use fixed parameter settings, making it impossible to track application effects in real time and make targeted adjustments, thus making it difficult for the models to maintain stable predictive performance over long-term use.

[0006] The output of existing systems does not adequately align with clinical needs, often providing only simple probability values ​​or feature annotations and lacking standardized clinical decision support information. This necessitates extensive secondary interpretation and processing by physicians in practical applications, reducing diagnostic efficiency. Furthermore, the model optimization process lacks a dynamic feedback mechanism, making it impossible to adjust in real time based on actual application results, hindering the continuous improvement of the model's predictive performance. Summary of the Invention

[0007] The purpose of this invention is to provide an artificial intelligence-based nuclear medicine image prediction system to solve the problems mentioned in the background art.

[0008] To achieve the above objectives, the present invention provides an artificial intelligence-based nuclear medicine image prediction system, the system comprising:

[0009] The image dynamic acquisition layer acquires nuclear medicine image data based on a multimodal temporal sampling mechanism, applies an adaptive noise reduction network to the nuclear medicine image data, generates a dynamic functional metabolic map, and annotates the dynamic functional metabolic map with anomaly probability.

[0010] The deep feature extraction layer deploys a 3D convolution module based on the labeled abnormal probability regions. It expands the receptive field through dilated convolution to capture deep features of the image and generate a multi-scale semantic feature map.

[0011] A cross-modal temporal fusion layer establishes the spatiotemporal correlation between static structure and dynamic function, uses an attention mechanism to synchronize feature weights, and fuses dynamic functional metabolic map and multi-scale semantic feature map through gated recurrent unit to generate disease risk probability.

[0012] The intelligent prediction and decision-making layer transforms the probability of disease risk into clinical prediction results based on reinforcement learning algorithms, and outputs the clinical prediction results to the terminal device in real time through a standardized interface.

[0013] The dynamic feedback optimization layer tracks the application of clinical prediction results received by the terminal device in real time, calculates the deviation between the actual application effect and the preset effect threshold, generates the model prediction efficacy index, and dynamically optimizes the nuclear medicine image prediction model until the model prediction efficacy index reaches stability.

[0014] Preferably, the method for acquiring nuclear medicine image data based on a multimodal temporal sampling mechanism includes:

[0015] The nuclear medicine imaging prediction system deploys three imaging acquisition devices: a PET scanner, a SPECT imager, and an MRI machine, to acquire multimodal imaging data of human organs. The multimodal imaging data includes organ metabolic activity data, blood perfusion data, and anatomical structure data.

[0016] A dynamic sampling mechanism simulating the physiological activity cycle dynamically adjusts the scanning frame rate and image resolution; based on human anatomical atlas, functionally sensitive areas, high-incidence lesion areas, and structural boundary areas of organs are designated as high-priority sampling areas; the remaining areas of organs are designated as regular sampling areas; a sampling weight function is dynamically defined to divide organ regions into high-priority sampling areas and regular sampling areas; high-priority sampling areas use high-density temporal sampling, while regular sampling areas use sparse temporal sampling.

[0017] The dynamic change characteristics in multimodal image data are analyzed in real time using a temporal difference network. The input of the temporal difference network is multimodal image data, and the output is the dynamic change region. The sampling area is dynamically adjusted according to the dynamic change region. If a new dynamic change region is detected, it is included in the high-priority sampling area. If no dynamic change is detected in the high-priority sampling area, it is downgraded to a regular sampling area. The high-priority sampling area adopts a continuous frame scanning mode, and the regular sampling area adopts an interval frame scanning mode. The acquisition device parameters are dynamically adjusted by control commands to obtain the final multimodal image data.

[0018] Preferably, the method for generating a dynamic functional metabolic map by applying an adaptive denoising network to nuclear medicine image data includes:

[0019] Intensity normalization is performed on nuclear medicine image data. The normalized nuclear medicine image data is then stacked into a five-dimensional tensor according to the modality dimension. The dimensions of the five-dimensional tensor include height, width, depth, number of modalities, and time frame. An adaptive denoising network structure is constructed, which includes a noise estimation module, a feature enhancement module, and a map reconstruction module.

[0020] Nuclear medicine image data is input into an adaptive denoising network structure. In the noise estimation module, a variational autoencoder is used to model the noise distribution of the input image data, generating a noise probability map. Noise features of different modalities are separated layer by layer, including PET noise features, SPECT noise features, and MRI noise features. In the feature enhancement module, the denoised image features are gradually restored to detail information through residual connections. A feedback connection is established between the noise estimation module and the feature enhancement module, feeding the noise probability map back to the feature enhancement module for targeted denoising.

[0021] 3D deconvolution is applied to the output of the adaptive denoising network structure to restore resolution. The multimodal denoising features output by the adaptive denoising network structure are spatiotemporally fused to generate a dynamic functional metabolic map.

[0022] Preferably, the method for annotating the anomaly probability of the dynamic functional metabolic profile includes:

[0023] The metabolic activity index of each voxel in the dynamic functional metabolic profile was calculated by weighted fusion of organ metabolic activity data, blood perfusion data and anatomical structure data included in nuclear medicine imaging data.

[0024] A first threshold and a second threshold for metabolic activity index are preset. The metabolic activity index is compared with the first threshold and the second threshold, respectively. If the metabolic activity index is less than the first threshold, the voxel corresponding to the index is marked as low abnormal probability; if the metabolic activity index is greater than the first threshold and less than the second threshold, the voxel corresponding to the index is marked as medium abnormal probability; if the metabolic activity index is greater than the second threshold, the voxel corresponding to the index is marked as high abnormal probability.

[0025] Different values ​​were used to label the abnormality probability of the dynamic functional metabolic profile. Different values ​​represent different abnormality probability levels. The low abnormality probability value was defined as 0.2, the medium abnormality probability value was defined as 0.5, and the high abnormality probability value was defined as 0.8.

[0026] Preferably, the method for generating multi-scale semantic feature maps includes:

[0027] Based on the labeled anomalous probability regions, feature extraction points of different densities are assigned to the dynamic functional metabolic map. These feature extraction points are located using 3D coordinates, and differentiated convolutional kernel sizes are applied to different anomalous probability regions. A 3D convolutional module is used to extract features from the image data, performing initial global feature mapping. The convolutional stride is dynamically adjusted according to the anomalous probability regions of the dynamic functional metabolic map, with the stride for high-probability regions being smaller than that for medium-probability regions, and vice versa. A multi-scale fusion algorithm is applied to dynamically integrate the convolutional features, introducing an attention weight mechanism. An iterative feature optimization algorithm continuously corrects the weight parameters in the feature map, stopping when a preset number of iterations is reached. A semantic feature hierarchy is established based on a feature pyramid model, and features at different scales are fused through adaptive pooling operations, ultimately generating a multi-scale semantic feature map with probability annotations.

[0028] Preferably, the method for establishing the spatiotemporal correlation between static structure and dynamic function includes:

[0029] Using the origin of the human anatomical coordinate system as a reference point, with the X and Y axes parallel to the cross-section and the Z axis perpendicular to the cross-section, a unified image coordinate system is defined. A calibration phantom is used to calibrate the image acquisition device, obtaining its spatial transformation parameters. The voxel coordinates of the dynamic functional metabolic atlas are transformed to the image coordinate system using these spatial transformation parameters. Then, anatomical atlas registration is used to transform the image coordinate system to the standard anatomical coordinate system, thereby obtaining the dynamic functional metabolic atlas in the standard anatomical coordinate system. Using the origin of the standard anatomical coordinate system as a reference point, coordinate transformation is performed on the multi-scale semantic feature atlas to obtain its spatial position parameters.

[0030] Multi-scale semantic feature maps are mapped to a standard anatomical coordinate system using spatial location parameters, thereby obtaining multi-scale semantic feature maps in the standard anatomical coordinate system. In the standard anatomical coordinate system, the dynamic functional metabolic map and the multi-scale semantic feature map are spatially grid-aligned to establish a spatial association between static structure and dynamic function. The timestamp of the dynamic functional metabolic map is set as the reference time, and the timestamp of the multi-scale semantic feature map is set as the synchronization time. Time calibration is performed using a time interpolation algorithm to establish a time association between static structure and dynamic function.

[0031] Preferably, the method for fusing dynamic functional metabolic maps and multi-scale semantic feature maps through gated recurrent units includes:

[0032] Based on the dynamic functional metabolic map and multi-scale semantic feature map in the standard anatomical coordinate system, a temporal feature sequence is constructed. A temporal fusion network is built using a deep learning structure based on bidirectional gated recurrent units to fuse the dynamic functional metabolic map and multi-scale semantic feature map in the standard anatomical coordinate system. The temporal fusion network includes an input layer, a feature alignment layer, a gated recurrent layer, a feature interaction layer, and an output layer. The temporal feature sequence is used as the input to the input layer of the temporal fusion network, and the output layer generates the lesion risk probability.

[0033] Preferably, the method for constructing the temporal feature sequence includes:

[0034] Each time frame in the dynamic functional metabolic atlas under the standard anatomical coordinate system is taken as a functional feature vector, and the dynamic change parameters of each functional feature vector are extracted as temporal functional features; each scale layer of the multi-scale semantic feature atlas under the standard anatomical coordinate system is taken as a structural feature vector, and the spatial distribution parameters of each structural feature vector are extracted as multi-scale structural features; all temporal functional features and multi-scale structural features are collected to obtain a feature set;

[0035] Iterate through all temporal functional features and calculate the cosine similarity of feature vectors of adjacent time frames; preset a time similarity threshold; if the cosine similarity of feature vectors of adjacent time frames is less than the preset time similarity threshold, then add a transition feature vector between the corresponding two time frames; if the cosine similarity of feature vectors of adjacent time frames is greater than or equal to the preset time similarity threshold, then do not add one.

[0036] Iterate through all multi-scale structural features and calculate the Euclidean distance between feature vectors at different scales. A preset scale distance threshold is used; if the Euclidean distance between feature vectors at different scales is less than the threshold, a fused feature vector is added between the corresponding two scales. If the Euclidean distance is greater than or equal to the threshold, no fused feature vector is added. Using a temporal alignment algorithm, temporal functional features and multi-scale structural features are arranged in chronological order, and correlation weights between features are added to obtain a sequence set. A temporal feature sequence is constructed based on the obtained feature set and sequence set.

[0037] Preferably, the method for converting lesion risk probability into clinical prediction results based on reinforcement learning algorithms includes:

[0038] Human organs are divided into multiple functional sub-regions, and each sub-region records the current disease risk probability and clinical relevance weight; at the same time, all possible clinical prediction indicators are listed, including the normal reference range, clinical significance level and historical prediction accuracy of each indicator.

[0039] Three predictive objectives were established, including lesion localization, progression trend, and intervention recommendation objectives; two types of constraints were set, including medical ethics constraints and data reliability constraints.

[0040] Using a reinforcement learning algorithm, n sets of clinical prediction indicator combinations are randomly generated. The achievement of the three prediction objectives is evaluated for each combination, and the achievement scores of the three prediction objectives are obtained. The achievement scores of the three prediction objectives are then weighted to obtain a comprehensive prediction score. The combination with the highest comprehensive prediction score is selected from the n sets of combinations as the final prediction combination. The selected final prediction combination is then transformed into clinically usable prediction results. The clinical prediction results include lesion location results, risk level results, and follow-up recommendations.

[0041] The methods for obtaining scores on the achievement of the three predicted objectives include:

[0042] The achievement of the lesion localization target is scored by comparing the overlap between the predicted lesion area and the actual lesion area; the achievement of the progression trend target is scored by calculating the deviation rate between the predicted progression rate and the clinical follow-up data; and the achievement of the intervention recommendation target is scored by statistically analyzing the proportion of intervention recommendations adopted clinically.

[0043] Preferably, the method for predicting the performance index using the generative model includes:

[0044] Collect the clinical application data fed back by the acquisition terminal device, synchronously evaluate the clinical compliance of the prediction results, calculate the deviation between the clinical compliance and the preset compliance threshold, and generate the 0-1 standardized model prediction efficacy index e; preset the first threshold e1 and the second threshold e2 of the standardized model prediction efficacy index; if e>e2, it is determined that the prediction model is effective; if e1<e<e2, it is determined that the prediction model needs parameter optimization; if e<e1, the prediction model is invalid, retrain the nuclear medicine image prediction model, and immediately trigger the model update mechanism.

[0045] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0046] This system obtains nuclear medicine image data through a multi-modal time-series sampling mechanism, applies an adaptive noise reduction network to generate a dynamic functional metabolism map, and combines a three-dimensional convolutional module and dilated convolution technology to capture deep image features, effectively solving the problems of incomplete single-modal data information and insufficient depth of feature extraction in traditional image analysis. The dynamic functional metabolism map can reflect the metabolic change process of tissues and organs in real time, while the multi-scale semantic feature map covers structural features at different levels. The combination of the two makes the system more comprehensive in identifying lesions and reduces misjudgments caused by information loss.

[0047] The cross-modal time-series fusion layer establishes the spatio-temporal correlation between static structures and dynamic functions, synchronizes feature weights using an attention mechanism, and realizes the deep fusion of dynamic and static features through a gated recurrent unit, overcoming the defects of separate processing of static and dynamic features and insufficient correlation modeling in the prior art. The attention mechanism can automatically focus on the feature regions more critical for lesion prediction, improving the pertinence of feature fusion; the gated recurrent unit effectively captures the temporal dependence relationship of dynamic data, enabling the fused features to more accurately reflect the development trend of lesions, thereby improving the accuracy of calculating the lesion risk probability.

[0048] The intelligent prediction and decision-making layer converts the lesion risk probability into a clinical prediction result based on a reinforcement learning algorithm and outputs it to the terminal device through a standardized interface, changing the situation of low compliance between the output result of the traditional system and clinical needs. The reinforcement learning algorithm can simulate the clinical decision-making process to make the output result more in line with the diagnosis logic of physicians; the standardized interface ensures the rapid transmission and direct application of the results, reducing the secondary processing work of physicians and improving the efficiency of clinical diagnosis.

[0049] The dynamic feedback optimization layer tracks the application of clinical prediction results in real time, calculates the deviation between the actual effect and the preset threshold, generates a model prediction efficacy index, and dynamically optimizes the model. This solves the problem of existing models lacking adaptive adjustment mechanisms and struggling to continuously improve predictive efficacy. This mechanism enables the system to continuously adjust model parameters based on actual clinical data, adapting to the characteristics and disease manifestations of different patient groups. This ensures that the model maintains stable predictive efficacy during long-term use and avoids efficacy decline due to model stagnation.

[0050] The collaborative work between the various levels of the system forms a closed-loop intelligent processing flow. From data acquisition, feature extraction, fusion analysis to prediction output and model optimization, each step addresses the shortcomings of existing technologies. Spatiotemporal correlation modeling of dynamic functions and static structures improves the accuracy of lesion identification; the combination of attention mechanisms and gated recurrent units enhances the effectiveness of feature fusion; the application of reinforcement learning algorithms improves the practicality of clinical prediction results; and dynamic feedback optimization ensures the long-term stability of the system. These improvements work together to enable the system to more accurately identify lesions and more efficiently assist in diagnostic decision-making in clinical applications, continuously improving performance as applications deepen, meeting the higher demands of nuclear medicine image analysis in terms of accuracy, efficiency, and adaptability. Attached Figure Description

[0051] Figure 1 This is a schematic diagram illustrating the working principle of the artificial intelligence-based nuclear medicine image prediction system described in this invention.

[0052] Figure 2 This is a flowchart for acquiring nuclear medicine image data based on a multimodal temporal sampling mechanism;

[0053] Figure 3 A flowchart for applying an adaptive noise reduction network to nuclear medicine imaging data;

[0054] Figure 4 This is a flowchart illustrating the fusion of dynamic functional metabolic maps and multi-scale semantic feature maps via a gated recurrent unit. Detailed Implementation

[0055] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0056] Please see Figures 1-4This invention provides an artificial intelligence-based nuclear medicine image prediction system. The system includes: a dynamic image acquisition layer, a deep feature extraction layer, a cross-modal temporal fusion layer, an intelligent prediction decision layer, and a dynamic feedback optimization layer. These layers work collaboratively to achieve intelligent prediction of nuclear medicine images. The specific implementation steps are as follows:

[0057] The image dynamic acquisition layer acquires nuclear medicine image data based on a multimodal temporal sampling mechanism. By deploying various image acquisition devices and combining dynamic sampling strategies, it selectively acquires image information from different regions and applies an adaptive noise reduction network to process the acquired image data, generating a dynamic functional metabolic map. At the same time, it annotates the map with anomaly probabilities, providing an accurate image basis for subsequent feature extraction.

[0058] The deep feature extraction layer deploys a 3D convolutional module based on the abnormal probability regions marked on the dynamic functional metabolic map. It uses dilated convolution technology to expand the receptive field, thereby capturing deep features in the image. After multi-scale processing, it generates a multi-scale semantic feature map to further mine the potential information in the image.

[0059] The cross-modal temporal fusion layer is responsible for establishing the spatiotemporal relationship between static structure and dynamic function. It adjusts feature weights with the help of attention mechanism, and then fuses dynamic functional metabolic map and multi-scale semantic feature map through gated recurrent unit to calculate the probability of disease risk, thereby achieving effective integration of different types of features.

[0060] The intelligent prediction decision layer is based on reinforcement learning algorithms, which transforms the probability of disease risk into specific clinical prediction results and outputs the results to terminal devices in real time through standardized interfaces, providing direct reference for clinical diagnosis.

[0061] The dynamic feedback optimization layer tracks the application of clinical prediction results by terminal devices in real time, calculates the deviation between the actual effect and the preset threshold, generates a model prediction efficacy index, and dynamically optimizes the nuclear medicine image prediction model based on the index until the model prediction efficacy reaches a stable state, ensuring that the system continues to maintain good prediction performance.

[0062] Example 1:

[0063] When acquiring nuclear medicine imaging data based on a multimodal temporal sampling mechanism, a PET scanner, a SPECT imager, and an MRI device are simultaneously deployed in the nuclear medicine imaging prediction system. The PET scanner captures metabolic activity data of human organs by detecting the distribution of positron-emitting radionuclide-labeled tracers, reflecting the functional metabolic state of organ tissues. The SPECT imager uses the distribution of radionuclides in the body to collect blood perfusion data of organs, presenting the flow and perfusion characteristics of blood in tissues. The MRI device, based on the principle of magnetic resonance, generates high-resolution anatomical data of human organs, clearly displaying the morphology and structural details of organs. These three imaging devices work collaboratively to acquire multimodal imaging data of human organs from different dimensions, forming a complementary information set.

[0064] A dynamic sampling mechanism is implemented to simulate the physiological activity cycle of the human body, dynamically adjusting the scanning frame rate and image resolution according to the physiological activity patterns of different organs. For example, for organs with strong periodic activity, such as the heart, different scanning frame rates are used during their systolic and diastolic phases to capture dynamic changes; for relatively stable organs, such as the liver, the scanning frame rate is appropriately reduced to minimize data redundancy. Based on human anatomical atlases, combined with clinical experience and pathological research results, functionally sensitive areas of organs (such as the hippocampus of the brain), areas with high incidence of lesions (such as the upper lobe of the lungs), and structural boundary areas (such as the junction between the organ and surrounding tissues) are designated as high-priority sampling areas; other areas of the organ, i.e., those with relatively stable function and low incidence of lesions, are designated as routine sampling areas. A sampling weight function is dynamically defined to quantify and distinguish the sampling priority of different regions of the organ, clearly defining the range and boundaries of high-priority sampling areas and routine sampling areas. High-priority sampling areas employ high-density temporal sampling, which involves performing more scans per unit time to obtain continuous and dense temporal image data; regular sampling areas employ sparse temporal sampling, which reduces the number of scans per unit time, thereby reducing the amount of data while preserving basic information.

[0065] A temporal difference network continuously runs to analyze dynamic changes in multimodal image data in real time. Its input is multimodal image data acquired by PET, SPECT, and MRI equipment. After processing by an internal algorithm, it outputs dynamic change regions in the images. Dynamic change regions include areas of sudden increase or decrease in metabolic activity, areas of abnormal fluctuation in blood perfusion, and areas of anatomical deformation. Based on the detection results of dynamic change regions, the sampling area is dynamically adjusted. If a new dynamic change region is detected, regardless of whether it originally belonged to a high-priority sampling area or a regular sampling area, it is included in the high-priority sampling area. If a previously high-priority sampling area shows no dynamic changes over multiple consecutive scan cycles, it is downgraded to a regular sampling area. The high-priority sampling area uses a continuous frame scan mode, meaning image acquisition is performed without interruption to ensure that dynamic changes at every time point are captured. The regular sampling area uses an interval frame scan mode, meaning scanning is performed at regular time intervals to balance data acquisition efficiency and information integrity. Through internal control commands, the parameters of the acquisition device are dynamically adjusted, including but not limited to scan frame rate, image resolution, scan range and radiation dose, so that the acquisition device can adapt in real time according to changes in the sampling area, and finally obtain optimized multimodal image data.

[0066] When generating dynamic functional metabolic atlases using an adaptive denoising network on nuclear medicine image data, the first step is to normalize the intensity of the acquired multimodal image data. By mapping the pixel values ​​of image data acquired from different modalities and time points to a unified numerical range, intensity deviations caused by differences in equipment and scanning parameters are eliminated, making the multimodal image data comparable. The normalized nuclear medicine image data are then stacked according to modal dimensions to form a five-dimensional tensor. The dimensions of this tensor are height, width, depth, modality number, and time frame. Height, width, and depth correspond to the spatial dimensions of the image, the modality number corresponds to the three different image types (PET, SPECT, and MRI), and the time frame corresponds to the image sequence acquired at different time points. The constructed adaptive denoising network structure includes a noise estimation module, a feature enhancement module, and an atlas reconstruction module, which are sequentially connected and work collaboratively.

[0067] Nuclear medicine image data is input into an adaptive denoising network structure. In the noise estimation module, a variational autoencoder analyzes the input image data layer by layer, constructing a probabilistic model to model the noise distribution in the image and generate a noise probability map. This noise probability map can intuitively display the noise intensity and distribution at different locations in the image. Simultaneously, the variational autoencoder can also separate noise features of different modalities layer by layer. PET noise features mainly consist of statistical noise and scattering noise; SPECT noise features include photon statistical noise and attenuation correction error; and MRI noise features are mainly Gaussian noise and motion artifacts. In the feature enhancement module, residual connections are used to transfer image detail information that may be lost during denoising from the input layer to subsequent layers, gradually restoring the image's edge, texture, and other detailed features. A feedback connection is established between the noise estimation module and the feature enhancement module, feeding the noise probability map back to the feature enhancement module in real time. This allows the feature enhancement module to adjust the denoising strategy according to the noise distribution and intensity, performing deeper denoising processing on areas with concentrated noise and retaining more original information in areas with less noise.

[0068] At the output of the adaptive denoising network structure, a 3D deconvolution operation is applied to restore the resolution of the processed image features. By progressively enlarging the size of the feature map, the spatial resolution of the image is improved, enabling the output image data to clearly display the fine structures of organs. The multimodal denoising features output by the adaptive denoising network structure are then spatiotemporally fused, combining the feature information of different modalities and the dynamic changes of different time frames to generate a dynamic functional metabolic atlas.

[0069] Example 2:

[0070] When annotating the probability of anomalies in dynamic functional metabolic maps, the metabolic activity index of each voxel in the map must first be calculated. This index is obtained by weighted fusion of organ metabolic activity data, blood perfusion data, and anatomical structure data contained in nuclear medicine imaging data. During the fusion process, each data type is assigned a corresponding weight according to its role in reflecting organ functional status. Organ metabolic activity data directly reflects the metabolic level of tissue cells, blood perfusion data reflects the blood supply to tissues, and anatomical structure data displays the morphological characteristics of tissues. All three participate in the calculation of the metabolic activity index to comprehensively reflect the physiological state of each voxel.

[0071] A first and a second threshold for the metabolic activity index are preset. These two thresholds, determined based on extensive clinical data and pathological studies, are used to distinguish different levels of abnormality probability. The metabolic activity index of each voxel is compared to these two thresholds. If the metabolic activity index is less than the first threshold, the voxel corresponding to that index is marked as having a low probability of abnormality; if the metabolic activity index is greater than the first threshold but less than the second threshold, it is marked as having a medium probability of abnormality; and if the metabolic activity index is greater than the second threshold, it is marked as having a high probability of abnormality. Different values ​​are used to label the abnormality probability of the dynamic functional metabolic profile, with a low probability of abnormality value set to 0.2, a medium probability of abnormality value set to 0.5, and a high probability of abnormality value set to 0.8. These values ​​allow for a direct differentiation of the abnormality probability levels of different voxels.

[0072] When generating a multi-scale semantic feature map, feature extraction points of varying densities are assigned to the map based on the anomalous probability regions marked on the dynamic functional metabolic map. These feature extraction points are located using three-dimensional coordinates, meaning their specific position in three-dimensional space is determined by the values ​​of the X, Y, and Z axes. Differentiated convolutional kernel sizes are used for different anomalous probability regions: smaller kernels are used for high anomalous probability regions to capture finer features; medium-sized kernels are used for medium-sized regions; and larger kernels are used for low-probability regions. A three-dimensional convolution module is used to extract features from the image data, performing initial global feature mapping to extract and present key features from the image as a whole. The convolutional stride is dynamically adjusted based on the anomalous probability regions of the dynamic functional metabolic map, with a smaller stride for high-probability regions than for medium-probability regions, and a smaller stride for medium-probability regions than for low-probability regions. This approach allows for denser feature extraction in high-probability regions and sparser feature extraction in low-probability regions.

[0073] A multi-scale fusion algorithm is applied to dynamically integrate convolutional features, fusing features extracted from different convolutional kernels and strides to form a more comprehensive feature set. An attention weighting mechanism is introduced, assigning different weights to different features, ensuring that important features occupy a larger proportion in the fusion process. An iterative feature optimization algorithm continuously corrects the weight parameters in the feature map; each iteration readjusts the weights based on the importance of the features. After multiple iterations, the weight parameters gradually become reasonable, and iteration stops when a preset number of iterations is reached. A semantic feature hierarchy structure is established based on a feature pyramid model. This structure corresponds to features of different scales from the bottom to the top, with bottom-level features containing more detailed information and top-level features containing more semantic information. Adaptive pooling operations are used to fuse features of different scales. The pooling window size and stride are automatically adjusted according to the feature scale and importance, organically combining features from different levels to ultimately generate a multi-scale semantic feature map with probability annotations. This map contains both anomaly probability information from different regions and integrates multi-scale semantic features, providing rich and targeted feature data for subsequent cross-modal temporal fusion.

[0074] Example 3:

[0075] When establishing the spatiotemporal relationship between static structure and dynamic function, the origin of the human anatomical coordinate system is used as the reference point. The X and Y axes are set parallel to the cross-section, and the Z axis is perpendicular to the cross-section, thus defining a unified image coordinate system. This coordinate system allows for standardized description of the spatial positions of different image data. A calibration phantom is used to calibrate the image acquisition device. The phantom contains markers of known size and location. By analyzing the imaging results of the acquisition device on these markers, the spatial transformation parameters of the acquisition device are calculated. These parameters include translation, rotation, and scaling factors, which are used to correct for deviations in spatial positioning between different devices.

[0076] The voxel coordinates of the dynamic functional metabolic atlas are transformed to the image coordinate system through spatial transformation parameters, ensuring that the spatial position of the dynamic functional metabolic atlas is consistent with the imaging coordinate system of the image acquisition device. Then, through anatomical atlas registration, the image coordinate system is transformed to the standard anatomical coordinate system. The standard anatomical coordinate system is established based on universal human anatomy standards, enabling spatial comparison between different individuals and different image data. After coordinate transformation, the dynamic functional metabolic atlas in the standard anatomical coordinate system is obtained. Using the origin of the standard anatomical coordinate system as the reference point, the same coordinate transformation is performed on the multi-scale semantic feature atlas, calculating the spatial position parameters of each feature point in the feature atlas in the standard anatomical coordinate system, including three-dimensional coordinate values ​​and spatial range.

[0077] Based on spatial location parameters, the multi-scale semantic feature atlas is mapped onto a standard anatomical coordinate system, placing it within the same spatial reference frame as the dynamic functional metabolic atlas. This allows for the acquisition of the multi-scale semantic feature atlas within the standard anatomical coordinate system. Within this system, both the dynamic functional metabolic atlas and the multi-scale semantic feature atlas are divided into spatial grids of equal size. Grid matching aligns the two atlases, ensuring precise correspondence between corresponding anatomical locations and establishing a spatial association between static structure and dynamic function. The timestamp of the dynamic functional metabolic atlas is set as the baseline time, recording the specific moment of image data acquisition. The timestamp of the multi-scale semantic feature atlas is set as the synchronization time. A temporal interpolation algorithm is used to calibrate the image data at different timestamps. When discrepancies exist between the two atlases' timestamps, feature data at intermediate moments is generated based on the characteristic change trends of adjacent time points, ensuring temporal synchronization and establishing a temporal association between static structure and dynamic function.

[0078] When fusing dynamic functional metabolic maps and multi-scale semantic feature maps using gated recurrent units, a temporal feature sequence is constructed based on these maps in a standard anatomical coordinate system. This temporal feature sequence contains functional and structural features from different time points. A deep learning architecture based on bidirectional gated recurrent units is used to build the temporal fusion network. The bidirectional gated recurrent units can simultaneously process forward and backward temporal information, capturing long-term dependencies in the temporal data.

[0079] The temporal fusion network comprises an input layer, a feature alignment layer, a gated recurrent layer, a feature interaction layer, and an output layer. The input layer receives the constructed temporal feature sequence and converts it into a tensor form that the network can process. The feature alignment layer further aligns the input features spatially and temporally, ensuring consistency between functional and structural features in the spatiotemporal dimensions. The gated recurrent layer consists of multiple bidirectional gated recurrent units, controlling the flow of feature information through a gating mechanism, retaining important temporal features, filtering redundant information, and learning temporal dependencies between features. The feature interaction layer interacts with the features output from the gated recurrent layer, achieving deep fusion of functional and structural features through feature matrix operations to generate a fused comprehensive feature. The output layer adopts a fully connected network structure, converting the comprehensive feature into lesion risk probabilities, which are represented numerically, reflecting the likelihood of lesions occurring in different regions. The temporal feature sequence is used as the input to the input layer of the temporal fusion network, processed sequentially through each layer, and finally output to generate the lesion risk probability.

[0080] Example 4:

[0081] When constructing the temporal feature sequence, each time frame in the dynamic functional metabolic atlas under the standard anatomical coordinate system is treated as a functional feature vector. Each functional feature vector contains information such as the metabolic activity index and dynamic change parameters of all voxels within that time frame. These parameters reflect the functional state of the organ at that time point, such as changes in metabolic rate and fluctuations in blood perfusion. The dynamic change parameters of each functional feature vector are extracted as temporal functional features, including the difference in metabolic activity between adjacent time frames, the rate of change, and the trend of change. Each scale layer of the multi-scale semantic feature atlas under the standard anatomical coordinate system is treated as a structural feature vector. Each structural feature vector encompasses the spatial distribution information, texture features, and association information with surrounding structures within that scale layer. The spatial distribution parameters of each structural feature vector are extracted as multi-scale structural features, including the three-dimensional coordinate distribution of feature points, region area, and morphological parameters. All temporal functional features and multi-scale structural features are collected to form a feature set containing multi-dimensional information in time and space.

[0082] Iterate through all temporal feature data and calculate the cosine similarity between feature vectors of adjacent time frames. Cosine similarity measures the similarity of two vectors in direction, and it is calculated by dividing the dot product of the two vectors by the product of their magnitudes, as shown in the following formula:

[0083]

[0084] in, The cosine similarity represents the feature vectors of adjacent time frames. This represents the i-th element of the feature vector of the previous time frame. This represents the i-th element of the feature vector of the next time frame, where n represents the dimension of the feature vector. A preset time similarity threshold is used, determined based on the range of feature changes under normal physiological conditions in clinical data. If the cosine similarity of the feature vectors of adjacent time frames is less than the preset time similarity threshold, it indicates a significant feature change between the two time frames. In this case, a transition feature vector is added between the two corresponding time frames. This transition feature vector is calculated by interpolating the feature vectors of the two adjacent time frames to smooth abrupt changes in the time series. If the cosine similarity of the feature vectors of adjacent time frames is greater than or equal to the preset time similarity threshold, it indicates a relatively gradual feature change between the two time frames, and no transition feature vector needs to be added.

[0085] The algorithm iterates through all multi-scale structural features, calculating the Euclidean distance between feature vectors at different scales. Euclidean distance reflects the straight-line distance between two vectors in space; by comparing the Euclidean distances of feature vectors at different scales, their spatial similarity can be determined. A preset scale distance threshold is established, based on the normal range of difference between features at different scales. If the Euclidean distance between feature vectors at different scales is less than the preset threshold, it indicates that the features at these two scales are spatially similar. A fused feature vector is added between the corresponding two scales, integrating the feature information from both scales to enhance the correlation between features at different scales. If the Euclidean distance between feature vectors at different scales is greater than or equal to the preset threshold, it indicates that the features at the two scales differ significantly, and no fused feature vector is needed. A temporal alignment algorithm is used to arrange temporal functional features and multi-scale structural features in chronological order, ensuring the feature sequence remains continuous in the time dimension. Simultaneously, correlation weights are added between features, determined by the magnitude of their correlation. Features with high correlation are assigned higher weights, and features with low correlation are assigned lower weights, thus forming a sequence set. Based on the acquired feature set and sequence set, the features are integrated according to time order and correlation to construct a complete temporal feature sequence. This sequence contains feature information extracted from multiple dimensions and time points, which can comprehensively reflect the functional and structural changes of organs.

[0086] Example 5:

[0087] When using reinforcement learning algorithms to convert disease risk probabilities into clinical predictions, human organs are first divided into multiple functional sub-regions. Taking the liver as an example, based on anatomical structure and functional characteristics, it can be divided into functional sub-regions such as the left lateral lobe, left medial lobe, right anterior lobe, right posterior lobe, and caudate lobe. Each sub-region independently records its current disease risk probability and clinical association weight. The clinical association weight is set according to the importance of the region in organ function and clinical pathogenesis characteristics. Simultaneously, all possible clinical prediction indicators are listed. For the lungs, these may include nodule size, metabolic activity value, edge morphology, and vascular convergence sign. Each indicator includes a normal reference range, clinical significance level, and historical prediction accuracy. For example, the normal reference range for nodule size may be set to a diameter of less than 5 mm. The clinical significance level is determined based on the degree of influence on disease assessment, and the historical prediction accuracy is derived from the statistical analysis of the correct proportion of disease predictions for that indicator in past cases.

[0088] Three predictive goals are established: lesion localization focuses on the specific location of the lesion within the organ; progression trend focuses on the changes in the lesion over time; and intervention recommendations focus on the clinical interventions to be taken in response to the lesion. Two types of constraints are set: medical ethics constraints cover the protection of patient privacy and recommendations to avoid unnecessary invasive examinations; and data reliability constraints require that all imaging data and clinical information used for prediction undergo quality verification to ensure traceability and compliance with collection standards.

[0089] When using reinforcement learning algorithms, n sets of clinical prediction indicator combinations are randomly generated. Taking the brain as an example, possible combinations include "metabolic activity value + blood perfusion data + anatomical structural abnormalities" and "metabolic activity value + temporal dynamic changes + edge morphology," among others. For each combination, the achievement of the three prediction goals is evaluated. A score for lesion localization is obtained by comparing the overlap between the predicted lesion area and the actual lesion area; higher overlap results in a higher score. A score for progression trend is obtained by calculating the deviation rate between the predicted progression rate and clinical follow-up data; lower deviation rate results in a higher score. A score for intervention recommendation achievement is obtained by statistically analyzing the proportion of intervention recommendations adopted clinically; higher adoption rates result in a higher score. The three scores are weighted to obtain a comprehensive prediction score, with weights determined based on the priority of the three goals in clinical decision-making. The combination with the highest comprehensive prediction score from the n sets of combinations is selected as the final prediction combination. The selected final prediction plan is transformed into clinically usable prediction results. For example, for lung lesions, the lesion location result may be clearly defined as the apical segment of the right upper lobe, the risk level result may be divided into three levels: low, medium and high, and the follow-up recommendation result may include the interval between re-examinations and the recommended imaging examination methods.

[0090] When generating the model prediction efficacy index, clinical application data feedback from the terminal device is collected. This data includes the usage of the prediction results by doctors, the diagnosis and treatment plans formulated based on the prediction results, and the subsequent clinical outcomes of patients. The clinical compliance of the prediction results is evaluated synchronously, that is, the degree of consistency between the prediction results and the actual clinical diagnosis results, such as the coincidence between the result predicted as a malignant tumor and the pathological examination result. Calculate the deviation between the clinical compliance and the preset compliance threshold, and generate a 0-1 standardized model prediction efficacy index e. The closer the e value is to 1, the better the model prediction efficacy. Preset the first threshold e1 and the second threshold e2 of the standardized model prediction efficacy index. Assume e1 is 0.6 and e2 is 0.8. If e > 0.8, it is judged that the prediction model is effective and can continue to be applied clinically; if 0.6 < e < 0.8, it is judged that the prediction model needs parameter optimization, such as adjusting the weights of feature extraction, optimizing the number of iterations of reinforcement learning, etc.; if e < 0.6, it is judged that the prediction model is ineffective, retrain the nuclear medicine image prediction model, train it using a new training data set and an adjusted network structure, and immediately trigger the model update mechanism to ensure that the terminal device uses the optimized model version. The whole process continues to cycle, and by continuously receiving clinical feedback data, the model is dynamically adjusted to make the prediction results more in line with the actual clinical needs.

[0091] It should be noted that in this article, relational terms such as first and second are only used to distinguish one entity or operation from another entity or operation, and do not necessarily require or imply any actual relationship or order between these entities or operations. Moreover, the terms "include", "comprise" or any other variant thereof are intended to cover non-exclusive inclusion, so that a process, method, article or device including a series of elements not only includes those elements, but also includes other elements not expressly listed, or also includes elements inherent to such process, method, article or device.

[0092] Although the embodiments of the present invention have been shown and described, those of ordinary skill in the art can understand that various changes, modifications, substitutions and variations can be made to these embodiments without departing from the principles and spirit of the present invention. The scope of the present invention is defined by the appended claims and their equivalents.

Claims

1. An artificial intelligence-based nuclear medicine image prediction system, characterized in that, include: The image dynamic acquisition layer acquires nuclear medicine image data based on a multimodal temporal sampling mechanism, applies an adaptive noise reduction network to the nuclear medicine image data, generates a dynamic functional metabolic map, and annotates the dynamic functional metabolic map with anomaly probability. The deep feature extraction layer deploys a 3D convolution module based on the labeled abnormal probability regions. It expands the receptive field through dilated convolution to capture deep features of the image and generate a multi-scale semantic feature map. A cross-modal temporal fusion layer establishes the spatiotemporal correlation between static structure and dynamic function, uses an attention mechanism to synchronize feature weights, and fuses dynamic functional metabolic map and multi-scale semantic feature map through gated recurrent unit to generate disease risk probability. The intelligent prediction and decision-making layer transforms the probability of disease risk into clinical prediction results based on reinforcement learning algorithms, and outputs the clinical prediction results to the terminal device in real time through a standardized interface. The dynamic feedback optimization layer tracks the application of clinical prediction results received by the terminal device in real time, calculates the deviation between the actual application effect and the preset effect threshold, generates the model prediction efficacy index, and dynamically optimizes the nuclear medicine image prediction model until the model prediction efficacy index reaches stability. The method for generating multi-scale semantic feature maps includes: Based on the labeled anomalous probability regions, feature extraction points of different densities are assigned to the dynamic functional metabolic map. These feature extraction points are located using 3D coordinates, and differentiated convolutional kernel sizes are applied to different anomalous probability regions. A 3D convolutional module is used to extract features from the image data, performing initial global feature mapping. The convolutional stride is dynamically adjusted according to the anomalous probability regions of the dynamic functional metabolic map, with the stride for high-probability regions being smaller than that for medium-probability regions, and vice versa. A multi-scale fusion algorithm is applied to dynamically integrate the convolutional features, introducing an attention weight mechanism. An iterative feature optimization algorithm continuously corrects the weight parameters in the feature map, stopping when a preset number of iterations is reached. A semantic feature hierarchy is established based on a feature pyramid model, and features at different scales are fused through adaptive pooling operations, ultimately generating a multi-scale semantic feature map with probability annotations.

2. The artificial intelligence-based nuclear medicine image prediction system according to claim 1, characterized in that, The method for acquiring nuclear medicine image data based on a multimodal temporal sampling mechanism includes: The nuclear medicine imaging prediction system deploys three imaging acquisition devices: a PET scanner, a SPECT imager, and an MRI machine, to acquire multimodal imaging data of human organs. The multimodal imaging data includes organ metabolic activity data, blood perfusion data, and anatomical structure data. A dynamic sampling mechanism simulating the physiological activity cycle dynamically adjusts the scanning frame rate and image resolution; based on human anatomical atlas, functionally sensitive areas, high-incidence lesion areas, and structural boundary areas of organs are designated as high-priority sampling areas; the remaining areas of organs are designated as regular sampling areas; a sampling weight function is dynamically defined to divide organ regions into high-priority sampling areas and regular sampling areas; high-priority sampling areas use high-density temporal sampling, while regular sampling areas use sparse temporal sampling. The dynamic change characteristics in multimodal image data are analyzed in real time using a temporal difference network. The input of the temporal difference network is multimodal image data, and the output is the dynamic change region. The sampling area is dynamically adjusted according to the dynamic change region. If a new dynamic change region is detected, it is included in the high-priority sampling area. If no dynamic change is detected in the high-priority sampling area, it is downgraded to a regular sampling area. The high-priority sampling area adopts a continuous frame scanning mode, and the regular sampling area adopts an interval frame scanning mode. The acquisition device parameters are dynamically adjusted by control commands to obtain the final multimodal image data.

3. The artificial intelligence-based nuclear medicine image prediction system according to claim 2, characterized in that, The method for generating dynamic functional metabolic maps by applying an adaptive noise reduction network to nuclear medicine imaging data includes: Intensity normalization is performed on nuclear medicine image data. The normalized nuclear medicine image data is then stacked into a five-dimensional tensor according to the modality dimension. The dimensions of the five-dimensional tensor include height, width, depth, number of modalities, and time frame. An adaptive denoising network structure is constructed, which includes a noise estimation module, a feature enhancement module, and a map reconstruction module. Nuclear medicine image data is input into an adaptive denoising network structure. In the noise estimation module, a variational autoencoder is used to model the noise distribution of the input image data, generating a noise probability map. Noise features of different modalities are separated layer by layer, including PET noise features, SPECT noise features, and MRI noise features. In the feature enhancement module, the denoised image features are gradually restored to detail information through residual connections. A feedback connection is established between the noise estimation module and the feature enhancement module, feeding the noise probability map back to the feature enhancement module for targeted denoising. 3D deconvolution is applied to the output of the adaptive denoising network structure to restore resolution. The multimodal denoising features output by the adaptive denoising network structure are spatiotemporally fused to generate a dynamic functional metabolic map.

4. The artificial intelligence-based nuclear medicine image prediction system according to claim 3, characterized in that, The method for annotating the anomaly probability of dynamic functional metabolic profiles includes: The metabolic activity index of each voxel in the dynamic functional metabolic profile was calculated by weighted fusion of organ metabolic activity data, blood perfusion data and anatomical structure data included in nuclear medicine imaging data. A first threshold and a second threshold for metabolic activity index are preset. The metabolic activity index is compared with the first threshold and the second threshold, respectively. If the metabolic activity index is less than the first threshold, the voxel corresponding to the index is marked as low abnormal probability; if the metabolic activity index is greater than the first threshold and less than the second threshold, the voxel corresponding to the index is marked as medium abnormal probability; if the metabolic activity index is greater than the second threshold, the voxel corresponding to the index is marked as high abnormal probability. Different values ​​were used to label the abnormality probability of the dynamic functional metabolic profile. Different values ​​represent different abnormality probability levels. The low abnormality probability value was defined as 0.2, the medium abnormality probability value was defined as 0.5, and the high abnormality probability value was defined as 0.

8.

5. The artificial intelligence-based nuclear medicine image prediction system according to claim 1, characterized in that, The method for establishing the spatiotemporal correlation between static structure and dynamic function includes: Using the origin of the human anatomical coordinate system as a reference point, with the X and Y axes parallel to the cross-section and the Z axis perpendicular to the cross-section, a unified image coordinate system is defined. A calibration phantom is used to calibrate the image acquisition device, obtaining its spatial transformation parameters. The voxel coordinates of the dynamic functional metabolic atlas are transformed to the image coordinate system using these spatial transformation parameters. Then, anatomical atlas registration is used to transform the image coordinate system to the standard anatomical coordinate system, thereby obtaining the dynamic functional metabolic atlas in the standard anatomical coordinate system. Using the origin of the standard anatomical coordinate system as a reference point, coordinate transformation is performed on the multi-scale semantic feature atlas to obtain its spatial position parameters. Multi-scale semantic feature maps are mapped to a standard anatomical coordinate system using spatial location parameters, thereby obtaining multi-scale semantic feature maps in the standard anatomical coordinate system. In the standard anatomical coordinate system, the dynamic functional metabolic map and the multi-scale semantic feature map are spatially grid-aligned to establish a spatial association between static structure and dynamic function. The timestamp of the dynamic functional metabolic map is set as the reference time, and the timestamp of the multi-scale semantic feature map is set as the synchronization time. Time calibration is performed using a time interpolation algorithm to establish a time association between static structure and dynamic function.

6. The artificial intelligence-based nuclear medicine image prediction system according to claim 5, characterized in that, The method for fusing dynamic functional metabolic maps and multi-scale semantic feature maps through gated recurrent units includes: Based on the dynamic functional metabolic map and multi-scale semantic feature map in the standard anatomical coordinate system, a temporal feature sequence is constructed. A temporal fusion network is built using a deep learning structure based on bidirectional gated recurrent units to fuse the dynamic functional metabolic map and multi-scale semantic feature map in the standard anatomical coordinate system. The temporal fusion network includes an input layer, a feature alignment layer, a gated recurrent layer, a feature interaction layer, and an output layer. The temporal feature sequence is used as the input to the input layer of the temporal fusion network, and the output layer generates the lesion risk probability.

7. The artificial intelligence-based nuclear medicine image prediction system according to claim 6, characterized in that, The method for constructing time-series feature sequences includes: Each time frame in the dynamic functional metabolic atlas under the standard anatomical coordinate system is taken as a functional feature vector, and the dynamic change parameters of each functional feature vector are extracted as temporal functional features; each scale layer of the multi-scale semantic feature atlas under the standard anatomical coordinate system is taken as a structural feature vector, and the spatial distribution parameters of each structural feature vector are extracted as multi-scale structural features; all temporal functional features and multi-scale structural features are collected to obtain a feature set; Iterate through all temporal functional features and calculate the cosine similarity of feature vectors of adjacent time frames; preset a time similarity threshold; if the cosine similarity of feature vectors of adjacent time frames is less than the preset time similarity threshold, then add a transition feature vector between the corresponding two time frames; if the cosine similarity of feature vectors of adjacent time frames is greater than or equal to the preset time similarity threshold, then do not add one. Traverse all multi-scale structural features, calculate the Euclidean distance of feature vectors at different scale levels; preset a scale distance threshold. If the Euclidean distance of feature vectors at different scale levels is less than the preset scale distance threshold, add a fused feature vector between the corresponding two scale levels; if the Euclidean distance of feature vectors at different scale levels is greater than or equal to the preset scale distance threshold, do not add; through the time series alignment algorithm, arrange the time series functional features and multi-scale structural features in chronological order, add the association weights between features, and obtain a sequence set; construct a time series feature sequence based on the obtained feature set and sequence set.

8. The artificial intelligence-based nuclear medicine image prediction system according to claim 7, characterized in that, The method for converting the lesion risk probability into a clinical prediction result based on the reinforcement learning algorithm includes: Divide the human organ into multiple functional sub-regions, and each sub-region records the current lesion risk probability and clinical association weight; list all possible clinical prediction indicators at the same time. The clinical prediction indicators include the normal reference range, clinical significance level, and historical prediction accuracy rate of each indicator. Establish three prediction goals, including the lesion localization goal, the progression trend goal, and the intervention suggestion goal; set two types of constraint conditions, including medical ethics constraints and data reliability constraints. Use the reinforcement learning algorithm to randomly generate n groups of clinical prediction indicator combination schemes, evaluate the achievement of the three prediction goals for each scheme, obtain the achievement scores of the three prediction goals, and perform weighted calculation on the achievement scores of the three prediction goals to obtain a comprehensive prediction score. Select the scheme with the highest comprehensive prediction score from the n groups of schemes as the final prediction scheme; convert the selected final prediction scheme into a clinically available prediction result; the clinical prediction result includes the lesion location result, the risk level result, and the follow-up suggestion result. The method for obtaining the achievement scores of the three prediction goals includes: By comparing the overlap degree between the predicted lesion area and the actual lesion area, it is used as the achievement score of the lesion localization goal; by calculating the deviation rate between the predicted progression speed and the clinical follow-up data, it is used as the achievement score of the progression trend goal; by counting the proportion of intervention suggestions adopted clinically, it is used as the achievement score of the intervention suggestion goal.

9. The artificial intelligence-based nuclear medicine image prediction system according to claim 8, characterized in that, The method for generating the prediction efficacy index of the generative model includes: Collect the clinical application data feedback by the terminal device, synchronously evaluate the clinical compliance of the prediction result, calculate the deviation between the clinical compliance and the preset compliance threshold, and generate a 0-1 standardized model prediction efficacy index e; preset a first threshold e1 and a second threshold e2 of the standardized model prediction efficacy index; if e > e2, it is judged that the prediction model is effective; if e1 < e < e2, it is judged that the prediction model needs parameter optimization; if e < e1, the prediction model is invalid, retrain the nuclear medicine image prediction model, and immediately trigger the model update mechanism.

Images