A Method for Temporal Variation Mining and Collaborative Analysis of Target Regions in Longitudinal CT Images

Abstract

This invention relates to the field of medical image processing technology, specifically to a method for mining and collaboratively analyzing the temporal changes of target regions in longitudinal CT images. The method includes the following steps: acquiring longitudinal CT image data and extracting features, outputting basic features at each time point; modeling the cross-temporal evolution evidence of the target region, outputting temporal evolution features; performing a change pattern analysis task, constructing a change pattern analysis loss; performing a trend analysis task, constructing a trend analysis loss; constructing a multi-task collaborative learning total loss including consistency constraints on change patterns and trends, completing joint training by minimizing the total loss, and outputting the temporal change analysis results of the target region. This invention can improve the consistency of feature representation across time points, while capturing short-term fluctuations and long-term evolution evidence, enhancing the completeness and stability of temporal change analysis, and improving the medical rationality and generalization ability of the model output through trend prior constraints.

Description

Technical Field

[0001] This invention relates to the field of medical image processing technology, specifically to a method for mining and collaborative analysis of temporal changes in target regions of longitudinal CT images. Background Technology

[0002] In longitudinal CT image processing, the same target region may exhibit multidimensional changes in overall volume, local morphology, edge structure, density distribution, and texture patterns at different follow-up time points. Compared with single-time-point CT images, longitudinal CT follow-up data can provide richer time-series information, helping to reveal the evolution of the target region in multiple follow-up stages. Therefore, it has important value in target region change analysis, temporal feature mining, and longitudinal image comparison.

[0003] However, the effective utilization of longitudinal CT information still faces several technical challenges:

[0004] (1) Evidence of changes across time points exhibits multi-scale characteristics. The target region may exhibit multiple change patterns, such as short-term slight fluctuations and long-term continuous evolution, at different follow-up intervals. If only a single time scale aggregation method is used, it is easy to miss one type of change evidence, thereby affecting the completeness of the temporal feature expression.

[0005] (2) Consistent representation of follow-up images. The scanning parameters, respiratory status of the subjects, differences in reconstruction algorithms, and phased appearance changes of the target area itself at different follow-up time points may lead to inconsistent feature distribution at different time points. If encoders with different parameters or inconsistent representation methods are used at each time point, representation bias is easily introduced, thereby affecting the stability of cross-time difference operations and temporal modeling.

[0006] (3) There is a coupling relationship between trend prior and change pattern analysis. The target region often shows a certain trend in longitudinal follow-up, such as local enhancement, basic stability, continuous weakening or stage fluctuation. If the trend information is only used as a simple additional input or coarse-grained auxiliary feature, it is easily affected by noise. A more reasonable approach is to train the trend analysis task and the change pattern analysis task together, and introduce the trend prior into the main task in the form of soft constraints through consistency constraints, thereby enhancing the stability and consistency of the time series analysis results.

[0007] (4) The trend labels have a blurred boundary problem. For trend categories such as shrinking, stabilizing, and enhancing, there are often blurred boundaries and transitional states in actual labeling. If multi-class learning is carried out directly, it is easy to bring about gradient instability and overfitting risks. Therefore, it is necessary to design more robust label representation and supervision methods.

[0008] Therefore, there is an urgent need for a technical solution that can consistently encode longitudinal CT images at multiple time points within a unified representation space, perform multi-scale mining and aggregation of temporal change evidence, and couple and constrain trend priors and change pattern analysis through multi-task collaboration, so as to improve the stability, completeness and reliability of temporal change analysis of target areas in longitudinal CT images. Summary of the Invention

[0009] To address the problems of insufficient utilization of longitudinal CT follow-up information, inconsistent cross-time characterization, easy loss of evidence of multi-scale changes, and difficulty in effectively constraining temporal change analysis by trend priors in existing technologies, this invention provides a method for mining and collaborative analysis of temporal changes in target regions of longitudinal CT images, specifically including the following steps: S1. Obtain longitudinal CT image data of the same target area of ​​the same subject at at least three follow-up time points, and extract features from the longitudinal CT image data based on a three-dimensional convolutional residual encoder with shared weights, and output the basic features at each time point.

[0010] S11. The number of follow-up time points is three, denoted as... , and And satisfy .

[0011] S12. At each time point, perform 3D region of interest cropping on the target area to obtain the corresponding 3D target area image data, denoted as... , and .

[0012] S13. The calculation formula for the basic feature is:

[0013] in Indicates basic features, Indicates a shared weight encoder. Indicates a time point index.

[0014] S2. The temporal multi-scale information mining module is used to model the cross-time evolution evidence of the target area and output the temporal evolution features.

[0015] S21. Perform time difference calculation on the basic characteristics at adjacent time points to obtain the change characteristics of adjacent follow-up stages. The calculation formula for the time difference operation is as follows:

[0016]

[0017] in, Indicates from a point in time At the appointed time The characteristics of change, Indicates from a point in time At the appointed time The characteristics of change, Indicates a point in time The basic characteristics Indicates a point in time The basic characteristics Indicates a point in time Its basic characteristics.

[0018] S22. Concatenate the basic features and the changing features in chronological order to construct a time-series feature sequence. The time-series feature sequence The formula is:

[0019] in, This indicates a splicing operation.

[0020] S23. For the time-series feature sequence Parallel multi-scale pooling is performed, and evolutionary evidence at different scales is weighted and aggregated through a self-attention mechanism to obtain temporal evolutionary features. :

[0021] in, This indicates a parallel multi-scale pooling module. This indicates a self-attention aggregation module.

[0022] S3. Perform a change pattern analysis task based on the aforementioned temporal evolution characteristics, and construct a change pattern analysis loss.

[0023] S31. Input the temporal evolution features into the first prediction head and output the analysis results of the change pattern of the target region. The results of the change pattern analysis The formula is:

[0024] in, Represents temporal evolution characteristics, and This represents the parameters of the change pattern analysis prediction head.

[0025] S32. Analysis results based on the aforementioned change pattern By combining the actual change pattern labels of the target area, a change pattern analysis loss is constructed. The change pattern analysis loss The formula is:

[0026] in, This represents the total number of change pattern categories. Labels indicating actual change patterns Indicates category index, Indicates that the prediction belongs to the first The probability of a change pattern. Indicates an indicator function, when Time indicator function The value is 1, otherwise it indicates the function. The value is 0.

[0027] S4. Perform a trend analysis task based on the time-series evolution characteristics and construct a trend analysis loss.

[0028] S41. Input the time-series evolution features into the second prediction head and output two trend prediction results. and The trend prediction results and The calculation formula is:

[0029]

[0030] in, Represents temporal evolution characteristics, , , and This represents the parameters of the trend prediction header.

[0031] S42, Define trend labels The value can be 0, 1, or 2, representing shrinkage, stabilization, and enhancement, respectively.

[0032] S43, Add trend tags Mapped to two binary supervised labels and The formula is:

[0033]

[0034] in, Indicates whether the target region is in a non-shrinkable state. Indicates whether the target area is in an enhanced state.

[0035] S44, Based on the binary supervision label , and trend forecast results , Build trend analysis loss The trend analysis loss The formula is:

[0036] in, This represents the binary cross-entropy loss function.

[0037] S5. Based on the change pattern analysis loss and trend analysis loss, construct a multi-task collaborative learning total loss including the consistency constraint of change pattern and trend coupling. Complete joint training by minimizing the total loss and output the temporal change analysis results of the target region.

[0038] S51, The formula for the coupling consistency constraint is:

[0039] in, This represents the loss value due to consistency constraints. This represents the consistency metric function. This represents a trend-inducing signal constructed from trend prediction probabilities. This indicates the results of the change pattern analysis.

[0040] S52, the total loss of the multi-task collaborative learning The formula is:

[0041] in, , and This represents the loss weighting coefficient. Indicates the loss from the change pattern analysis. Indicates the loss from trend analysis. This represents the loss value due to consistency constraints.

[0042] S53, By minimizing the total loss of the multi-task collaborative learning The shared weight encoder, the temporal multi-scale information mining module, the first prediction head and the second prediction head are jointly optimized and trained end-to-end, and the temporal change analysis results of the target region are output.

[0043] Compared with the prior art, the present invention has the following beneficial effects: (1) This invention is aimed at longitudinal CT follow-up scenarios. It uses a three-dimensional convolutional residual encoder with shared weights to consistently represent the target area images at different time points, so that the features at each time point are in a unified feature space, reducing the representation offset caused by differences in scanning parameters, reconstruction differences and stage appearance changes, thereby improving the stability of cross-time comparison and subsequent temporal modeling.

[0044] (2) This invention constructs a time series feature sequence of time point features and adjacent time difference features, and introduces a time series multi-scale information mining module to perform parallel convergence and attention fusion of evolution evidence at different time scales. It can simultaneously capture multi-scale change patterns such as short-term fluctuations, stage changes and long-term evolution, avoid the loss of key information caused by single time scale aggregation, and enhance the expressive power and change analysis capabilities of time series evolution features.

[0045] (3) The present invention adopts a multi-task collaborative learning framework, introduces a trend analysis task in addition to the change pattern analysis task, and decomposes the original trend supervision into two levels of binary supervision to reduce the training instability caused by the ambiguity of trend category boundaries, thereby guiding the main task to learn the longitudinal evolution law with more robust auxiliary supervision and improving the model's adaptability to noise, label uncertainty and distribution changes.

[0046] (4) The present invention further constructs a trend and change pattern coupling consistency constraint. By mapping the trend prediction result to the trend-induced change prior and constraining the change pattern analysis result, the longitudinal evolution trend prior of the target area is injected into the training process in the form of soft constraint, so that the model output is more in line with the temporal evolution law and the stability and credibility of the analysis result are improved. The overall structure consists of a shared encoder, a temporal multi-scale information mining module and two prediction heads, which is easy to perform end-to-end training and apply to longitudinal CT image processing. Attached Figure Description

[0047] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0048] Figure 1 This is a flowchart of a method for mining and collaborative analysis of temporal changes in target regions of longitudinal CT images according to the present invention.

[0049] Figure 2 This is a schematic diagram of the original ROI of a CT slice containing the target region in an embodiment of the present invention.

[0050] Figure 3This is a schematic diagram of the shared-weight 3D convolutional residual encoder structure in an embodiment of the present invention.

[0051] Figure 4 This is a schematic diagram of the temporal multi-scale information mining structure in an embodiment of the present invention.

[0052] Figure 5 This is a schematic diagram of the target region time-series change response ROI output in an embodiment of the present invention. Detailed Implementation

[0053] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. 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 should fall within the scope of protection of the present invention.

[0054] It should be noted that the terms "first," "second," etc., in the specification, claims, and accompanying drawings of this invention are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. It should be understood that such data can be interchanged where appropriate so that the embodiments of the invention described herein can be implemented in orders other than those illustrated or described herein. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover a non-exclusive inclusion; for example, a process, method, system, product, or apparatus that comprises a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to such processes, methods, products, or apparatus.

[0055] This invention provides the following technical solutions: like Figure 1 As shown, this invention discloses a method for mining and collaborative analysis of temporal changes in target regions of longitudinal CT images, mainly including the following steps: S1. Obtain longitudinal CT image data of the same target area of ​​the same subject at at least three follow-up time points, and extract features from the longitudinal CT image data based on a three-dimensional convolutional residual encoder with shared weights, and output the basic features at each time point.

[0056] S11. The number of follow-up time points is three, denoted as... , and And satisfy .

[0057] S12. At each time point, perform 3D region of interest cropping on the target area to obtain the corresponding 3D target area image data, denoted as... , and .

[0058] S13. The calculation formula for the basic feature is:

[0059] in Indicates basic features, Indicates a shared weight encoder. Indicates a time point index.

[0060] S2. The temporal multi-scale information mining module is used to model the cross-time evolution evidence of the target area and output the temporal evolution features.

[0061] S21. Perform time difference calculation on the basic characteristics at adjacent time points to obtain the change characteristics of adjacent follow-up stages. The calculation formula for the time difference operation is as follows:

[0062]

[0063] in, Indicates from a point in time At the appointed time The characteristics of change, Indicates from a point in time At the appointed time The characteristics of change, Indicates a point in time The basic characteristics Indicates a point in time The basic characteristics Indicates a point in time Its basic characteristics.

[0064] S22. Concatenate the basic features and the changing features in chronological order to construct a time-series feature sequence. The time-series feature sequence The formula is:

[0065] in, This indicates a splicing operation.

[0066] S23. For the time-series feature sequence Parallel multi-scale pooling is performed, and evolutionary evidence at different scales is weighted and aggregated through a self-attention mechanism to obtain temporal evolutionary features. :

[0067] in, This indicates a parallel multi-scale pooling module. This indicates a self-attention aggregation module.

[0068] S3. Perform a change pattern analysis task based on the aforementioned temporal evolution characteristics, and construct a change pattern analysis loss.

[0069] S31. Input the temporal evolution features into the first prediction head and output the analysis results of the change pattern of the target region. The results of the change pattern analysis The formula is:

[0070] in, Represents temporal evolution characteristics, and This represents the parameters of the change pattern analysis prediction head.

[0071] S32. Analysis results based on the aforementioned change pattern By combining the actual change pattern labels of the target area, a change pattern analysis loss is constructed. The change pattern analysis loss The formula is:

[0072] in, This represents the total number of change pattern categories. Labels indicating actual change patterns Indicates category index, Indicates that the prediction belongs to the first The probability of a change pattern. Indicates an indicator function, when Time indicator function The value is 1, otherwise it indicates the function. The value is 0.

[0073] S4. Perform a trend analysis task based on the time-series evolution characteristics and construct a trend analysis loss.

[0074] S41. Input the time-series evolution features into the second prediction head and output two trend prediction results. and The trend prediction results and The calculation formula is:

[0075]

[0076] in, Represents temporal evolution characteristics, , , and This represents the parameters of the trend prediction header.

[0077] S42, Define trend labels The value can be 0, 1, or 2, representing shrinkage, stabilization, and enhancement, respectively.

[0078] S43, Add trend tags Mapped to two binary supervised labels and The formula is:

[0079]

[0080] in, Indicates whether the target region is in a non-shrinkable state. Indicates whether the target area is in an enhanced state.

[0081] S44, Based on the binary supervision label , and trend forecast results , Build trend analysis loss The trend analysis loss The formula is:

[0082] in, This represents the binary cross-entropy loss function.

[0083] S5. Based on the change pattern analysis loss and trend analysis loss, construct a multi-task collaborative learning total loss including the consistency constraint of change pattern and trend coupling. Complete joint training by minimizing the total loss and output the temporal change analysis results of the target region.

[0084] S51, The formula for the coupling consistency constraint is:

[0085] in, This represents the loss value due to consistency constraints. This represents the consistency metric function. This represents a trend-inducing signal constructed from trend prediction probabilities. This indicates the results of the change pattern analysis.

[0086] S52, the total loss of the multi-task collaborative learning The formula is:

[0087] in, , and This represents the loss weighting coefficient. Indicates the loss from the change pattern analysis. Indicates the loss from trend analysis. This represents the loss value due to consistency constraints.

[0088] S53, By minimizing the total loss of the multi-task collaborative learning The shared weight encoder, the temporal multi-scale information mining module, the first prediction head and the second prediction head are jointly optimized and trained end-to-end, and the temporal change analysis results of the target region are output.

[0089] Example S1. Dataset processing and segmentation Longitudinal CT image data of the same target area of ​​the same subject at at least three follow-up time points are acquired. A three-dimensional convolutional residual encoder based on shared weights is used to extract features from the longitudinal CT image data and output the basic features at each time point.

[0090] A target region dataset was constructed by collecting longitudinal CT follow-up data from public datasets and hospital-owned private datasets. Each sample corresponds to CT scans of the same target region at three follow-up time points and has a change pattern label and a trend label. The change pattern label describes the temporal change status of the target region during the longitudinal follow-up process, and the trend label describes the overall change direction of the target region on the longitudinal time axis.

[0091] For each time point, CT images are cropped in 3D based on the target region's location to obtain ROI volume data containing the target region. For example... Figure 2 The diagram shown is a schematic representation of the original ROI of a CT slice containing the target region in this embodiment. There are three follow-up time points, denoted as... , and And satisfy To ensure consistency in model input, the spatial dimensions of the ROI are uniformly set to the preset voxel dimensions after cropping, for example... Furthermore, window width and level limits and normalization are applied to the intensity data to ensure consistency in intensity range across different data sources, reducing the adverse effects of intensity distribution differences on training. Taking min-max normalization as an example, its expression is:

[0092] in For voxel strength, The minimum value after truncation. This is the maximum value after truncation.

[0093] This invention employs 5-fold cross-validation, which requires ensuring that samples of the same tested object do not appear in the training set and test set simultaneously to avoid data leakage.

[0094] Let the change mode label be denoted as Trend tags are denoted as ,in These represent shrinkage, stabilization, and enhancement, respectively. To facilitate training for the trend analysis task, the trend labels are mapped to two levels of binary supervision labels, expressed by the following formula:

[0095] in, Indicates whether the target region is in a non-shrinkable state. Indicates whether the target area is in an enhanced state.

[0096] S2, Basic Feature Extraction Stage Using a Weighted Shared Encoder like Figure 3 The diagram shows the shared-weight 3D convolutional residual encoder structure in this embodiment. The 3D convolutional residual encoder, built using the concept of 3D-ResNet, encodes the ROI at three time points and outputs the basic features at those three time points. Shared weights mean that the same set of parameters acts on the inputs at the three time points, ensuring that the features at each time point reside in a unified representation space. This facilitates subsequent cross-temporal difference operations and temporal variation modeling. The feature extraction process is represented as follows:

[0097] The encoder consists of multi-level residual blocks with the number of channels progressively increasing, for example, 16, 32, 64, and 128, to achieve a three-dimensional structural representation from low-level texture to high-level semantics; residual connections are used to stabilize gradient propagation and improve training convergence.

[0098] S3. Temporal multi-scale information mining stage of features like Figure 4 As shown on the left, the basic features at adjacent time points are first differentiated to obtain two change features, which respectively characterize the changes from... and Evidence of changes. Subsequently, the time-point features and difference features are concatenated in chronological order to form a time-series feature sequence. .in, Indicates from a point in time At the appointed time The characteristics of change, Indicates from a point in time At the appointed time The characteristics of change, Indicates a point in time The basic characteristics Indicates a point in time The basic characteristics Indicates a point in time Its basic characteristics.

[0099] like Figure 4 As shown on the right, for Pooling or convergence operations of different scales are applied in parallel to capture evolutionary patterns at different time scales, denoted as... , , Smaller scales are more sensitive to short-term fluctuations, while larger scales are more biased towards overall trends. The outputs at each scale are then reshaped to form a sequence suitable for attention mechanisms, resulting in a multi-scale time series representation.

[0100] The multi-scale sequence representation is input into a multi-head attention module to learn the correlation and importance weights between evidence at different scales, and then pooling is used to obtain the temporal evolution features. This feature serves as a common input representation for subsequent multi-task collaborative losses. In this way, the present invention can simultaneously cover short-term changes, phased changes, and long-term evolutionary evidence within a unified temporal representation framework, thereby enhancing the expressive power of temporal change features.

[0101] S4. Employ multi-task collaborative loss to complete the target area temporal change discrimination stage. Temporal evolution characteristics The input is fed into the change pattern analysis and prediction head, which outputs the change pattern probability distribution or change score result. This, along with the change pattern label, constitutes the main task supervision, directly serving the analysis output of the temporal change status of the target area. The expression for the change pattern analysis result is:

[0102] in, This is the result of the change pattern analysis.

[0103] The temporal evolution features are simultaneously input into the trend prediction head, outputting two trend probabilities. The three-class trend supervision is converted into two levels of binary supervision: "whether to non-shrink" and "whether to enhance," to address the training instability caused by blurred class boundaries when directly performing multi-class supervision, and to improve the learnability and stability of the trend analysis task. The expression for the trend analysis result is as follows:

[0104] This embodiment further introduces trend-change pattern coupling consistency constraints in multi-task training: that is, using the trend prediction results to construct trend-induced change priors, and applying consistency constraints to them and change pattern analysis results, so that the model is guided by the soft constraints of trend priors while learning change pattern analysis tasks, thereby improving the consistency and generalization ability of time series analysis results to a certain extent.

[0105] The formula for the total loss is expressed as follows:

[0106] After training, inference is performed on the input ROI, and the output results include the change pattern of the target region, change score, or time-series change analysis, and the following are generated: Figure 5 The temporal change response ROI visualization results shown are used for change analysis and result presentation of longitudinal CT images.

[0107] Based on the solution of this invention, the experimental analysis is as follows: (1) Experimental data processing and segmentation This embodiment conducts experiments on the aforementioned target region samples. A data partitioning strategy based on the tested object is adopted to ensure that there is no overlap between the same tested object in the training set and the test set; at the same time, 5-fold cross-validation is used for stability evaluation.

[0108] The model is implemented using the deep learning framework PyTorch and trained on a GPU. The optimizer is Adam, and the learning rate, batch size, and number of training epochs are set and adjusted according to hardware resources and data scale. To improve generalization ability, augmentation operations such as random flipping, random rotation, and random intensity perturbation are applied to the ROI during the training phase.

[0109] (2) Evaluation indicators After completing model training, predictions are made on the test set, and the following evaluation metrics are calculated. Let true positive be TP, false positive be FP, true negative be TN, and false negative be FN: Accuracy (ACC): Determines performance when correctly classifying classes in the test set.

[0110]

[0111] Area under the ROC curve (AUC): Plot the TPR / FPR at different thresholds and calculate the area.

[0112] Area under the PR curve (AUPRC): Plots the Precision / Recall ratio at different thresholds and calculates the area, where:

[0113] F1 score (F1): It takes into account both the precision and recall of the classification model and is the harmonic mean of precision and recall.

[0114]

[0115] Kappa coefficient (Kappa): Used to measure categorical consistency, calculated according to the statistical definition.

[0116] (3) Analysis of model results To verify the effectiveness of the method of the present invention, it was compared with existing longitudinal CT time series analysis methods. The comparative experimental results are shown in Table 1 below: Table 1 Comparative Experimental Results

[0117] In Table 1, bold results represent the best results among the relevant indicators, and underlined results represent the second best results.

[0118] The comparative experiments are expected to demonstrate that this invention constructs cross-temporal difference and time-series sequences within a unified shared coding space, and simultaneously covers short-term and long-term change evidence through temporal multi-scale information mining and attention fusion. Furthermore, multi-task collaborative training uses trend analysis supervision as an auxiliary signal to guide the main task in learning longitudinal evolution patterns, and injects trend priors into the analysis process as soft constraints through coupling consistency constraints. Therefore, this invention is expected to have advantages in metrics reflecting overall analytical capability and imbalance robustness, such as AUC and AUPRC, while exhibiting more stable performance in comprehensive metrics such as F1 and Kappa.

[0119] To verify the necessity of each component module of this invention, this embodiment sets up three ablation experiments on the target region dataset: wo / TSMI: the temporal multi-scale information mining module is removed, and only the basic features are directly concatenated for task prediction; wo / Remove the trend prediction loss (at this point, the trend branch does not participate in training, and the co-loss is also cancelled), and only train for the benign / malignant discrimination task; wo / : Remove the collaborative loss, that is, remove the coupling consistency constraint term and only retain the cross-entropy supervision of each task; Ours: the complete model.

[0120] The results of the ablation experiment are shown in Table 2 below: Table 2 Ablation Experiment Results for Each Module

[0121] The results in the table show that: when the temporal multi-scale information mining module (wo / TSMI) is removed, the model can only rely on simple aggregation at a single scale, making it difficult to simultaneously capture complementary evidence of short-term fluctuations and long-term evolution. Consequently, the expected performance of AUC / AUPRC and other indicators is significantly reduced, indicating the importance of multi-scale mining and attention fusion for longitudinal evolution representation. When the co-operation loss (wo / CoopLoss) is removed, although the trend prediction task is still retained, the trend prior cannot feed back into the discrimination task in the form of consistency constraints. This makes it difficult for the discrimination branch to stably utilize trend information, resulting in reduced medical consistency of the prediction results and a decrease in the expected comprehensive indicators. This indicates that coupling consistency constraints contribute to improving stability. When the trend prediction loss (wo / TrendLoss) is removed, the model degenerates into single-task classification training using only longitudinal evolution features. It loses the auxiliary constraints and regularization effects brought by trend supervision, making it more prone to overfitting, especially when the sample size is limited or the classes are imbalanced. Consequently, the expected performance of AUPRC, F1, and Kappa indicators is reduced, indicating that multi-task co-operation learning can effectively improve generalization and robustness.

[0122] In summary, this embodiment verifies the feasibility and effectiveness of the present invention in the task of analyzing temporal changes in target areas during longitudinal CT follow-up. The above descriptions are merely preferred embodiments of the present invention. Any adjustments or substitutions made by those skilled in the art, such as encoder channel configuration, number of residual blocks, number of multi-scale branches, number of attention heads, loss weights, and data augmentation strategies, without departing from the concept of the present invention, should fall within the protection scope of the present invention.

[0123] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for mining and collaborative analysis of temporal changes in target regions of longitudinal CT images, characterized in that, Includes the following steps: Longitudinal CT image data of the same target area of ​​the same subject at at least three follow-up time points are acquired. A three-dimensional convolutional residual encoder based on shared weights is used to extract features from the longitudinal CT image data and output the basic features at each time point. A time-series multi-scale information mining module is used to model the cross-time evolution evidence of the target region and output time-series evolution features; Based on the aforementioned temporal evolution characteristics, a change pattern analysis task is performed, and a change pattern analysis loss is constructed. Based on the aforementioned temporal evolution characteristics, a trend analysis task is performed, and a trend analysis loss is constructed. Based on the change pattern analysis loss and trend analysis loss, a multi-task collaborative learning total loss including the consistency constraint of change pattern and trend coupling is constructed. Joint training is completed by minimizing the total loss, and the temporal change analysis results of the target region are output.

2. The method for temporal variation mining and collaborative analysis of target regions in longitudinal CT images according to claim 1, characterized in that, The steps for calculating the basic characteristics at each time point include: The number of follow-up time points is three, denoted as... , and And satisfy ; At each time point, a 3D region of interest (ROI) is cropped from the target area to obtain the corresponding 3D target area image data, denoted as . , and ; The formula for calculating the basic features is as follows: in Indicates basic features, Indicates a shared weight encoder. Indicates a point-in-time index.

3. The method for temporal variation mining and collaborative analysis of target regions in longitudinal CT images according to claim 1, characterized in that, The steps for modeling the cross-temporal evolution evidence of the target region and outputting temporal evolution features include: By performing time difference calculations on the basic characteristics at adjacent time points, the change characteristics of adjacent follow-up stages are obtained. The calculation formula for the time difference calculation is as follows: in, Indicates from a point in time At the appointed time The characteristics of change, Indicates from a point in time At the appointed time The characteristics of change, Indicates a point in time The basic characteristics Indicates a point in time The basic characteristics Indicates a point in time Basic characteristics; The basic features and the changing features are concatenated in chronological order to construct a time-series feature sequence. The time-series feature sequence The formula is: in, Indicates a splicing operation; For the time-series feature sequence Parallel multi-scale pooling is performed, and evolutionary evidence at different scales is weighted and aggregated through a self-attention mechanism to obtain temporal evolutionary features. : in, This indicates a parallel multi-scale pooling module. This indicates a self-attention aggregation module.

4. The method for temporal variation mining and collaborative analysis of target regions in longitudinal CT images according to claim 1, characterized in that, The steps for performing change pattern analysis based on the aforementioned temporal evolution characteristics and constructing the change pattern analysis loss include: The temporal evolution features are input into the first prediction head, and the analysis results of the change patterns in the target region are output. The results of the change pattern analysis The formula is: in, Represents temporal evolution characteristics, and Indicates the parameters of the change pattern analysis prediction head; Based on the analysis results of the aforementioned change patterns By combining the actual change pattern labels of the target area, a change pattern analysis loss is constructed. The change pattern analysis loss The formula is: in, This represents the total number of change pattern categories. Labels indicating actual change patterns Indicates category index, Indicates that the prediction belongs to the first The probability of a change pattern. Indicates an indicator function, when Time indicator function The value is 1, otherwise it indicates the function. The value is 0.

5. The method for temporal variation mining and collaborative analysis of target regions in longitudinal CT images according to claim 1, characterized in that, The steps of performing trend analysis based on the aforementioned temporal evolution characteristics and constructing the trend analysis loss include: The temporal evolution features are input into the second prediction head, and two trend prediction results are output. and The trend prediction results and The calculation formula is: in, Represents temporal evolution characteristics, , , and Indicates the trend prediction header parameters; Define trend labels The value can be 0, 1, or 2, representing shrinkage, stabilization, and enhancement, respectively. Trend tags Mapped to two binary supervised labels and The formula is: in, Indicates whether the target region is in a non-shrinkable state. Indicates whether the target area is in an enhanced state; Based on the binary supervision label , and trend forecast results , Build trend analysis loss The trend analysis loss The formula is: in, This represents the binary cross-entropy loss function.

6. The method for temporal variation mining and collaborative analysis of target regions in longitudinal CT images according to claim 1, characterized in that, The formula for the coupling consistency constraint is: in, This represents the loss value due to consistency constraints. This represents the consistency metric function. This represents a trend-inducing signal constructed from trend prediction probabilities. This indicates the results of the change pattern analysis.

7. The method for temporal variation mining and collaborative analysis of target regions in longitudinal CT images according to claim 1, characterized in that, The total loss of multi-task collaborative learning The formula is: in, , and This represents the loss weighting coefficient. Indicates the loss in the change pattern analysis. Indicates the loss from trend analysis. This represents the loss value due to consistency constraints.

8. The method for temporal variation mining and collaborative analysis of target regions in longitudinal CT images according to claim 7, characterized in that, By minimizing the total loss of the multi-task collaborative learning The shared weight encoder, the temporal multi-scale information mining module, the first prediction head and the second prediction head are jointly optimized and trained end-to-end, and the temporal change analysis results of the target region are output.

Images