Prospective classification device for predicting dementia and operation method of the same

Abstract

A prospective classification device for predicting dementia and a method of operation thereof are disclosed, which converts features of diagnostic brain imaging data acquired at the time of diagnosis of a patient with mild cognitive impairment by a prospective classification model into features of prognostic brain imaging data corresponding to a prognostic time after the time of diagnosis, and predicts the risk of a patient with mild cognitive impairment transitioning into a patient with dementia based on the features of prognostic brain imaging data converted from the features of diagnostic brain imaging data.

Description

Technology Field

[0001] The present invention relates to a prospective classification device for predicting dementia and a method of operating the same, and more specifically, to a prospective classification device for predicting the transition from mild cognitive impairment to Alzheimer's disease and a method of operating the same.

[0002] This invention is a technology developed through the Innovative Chronic Cerebrovascular Disease Biobank (Sub-project No.: KBN4-B02-2023-01) support project. Meanwhile, the Korea National Institute of Health, Korea Disease Control and Prevention Agency, has no property interest in any aspect of this invention. Background Technology

[0003] Alzheimer's disease (AD) is the most common form of dementia affecting the elderly. Driven by the rapid increase in the elderly population and the aging of society, the number of AD patients continues to rise; the global AD population is projected to triple from approximately 50 million in 2015 to 131.5 million by 2050. While Alzheimer's disease has recently emerged as a serious issue, prevention and the delay of progression remain the only solutions, as the cause of the disease is unclear and there is no cure. However, most patients undergo testing and receive a diagnosis after the disease has already developed. Therefore, it is crucial to detect potential risks early and predict disease progression.

[0004] While there is currently a tendency to define the transition to AD as a continuum of beta-amyloid accumulation, mild cognitive impairment (MCI) is considered a precursor to AD based on clinical symptoms. The slight memory or cognitive impairment experienced by MCI patients is regarded as an early stage of AD symptoms, such as long-term memory loss, language difficulties, disorientation, and personality changes. Previous studies have shown that approximately 12% of subjects with MCI progress to AD within four years of the onset of symptoms. Consequently, early diagnosis of potential AD risks leads to the task of predicting whether MCI patients will transition to AD.

[0005] MCI and AD are associated with changes observed in brain imaging, particularly brain volume loss (atrophy) and the occurrence of focal lesions in the white and gray matter. Advancements in medical imaging technologies, such as magnetic resonance imaging (MRI), have led to the development of various methods to enhance the capabilities of computer-aided systems that facilitate the early detection of AD. Furthermore, the application of machine learning algorithms has made it possible to more accurately detect subtle patterns of brain volume reduction. Generally, machine learning applied to brain MRI involves training algorithms on preprocessed feature sets, such as local volume and cortical thickness, to generate classifiers that predict correct diagnostic outcomes for new observations.

[0006] To predict the transition from MCI to AD, existing methods have been performed using supervised learning techniques aimed at deriving decision surfaces from image sets labeled MCI-C (transition from MCI to AD) and MCI-N (non-transition from MCI to AD). The most widely used classifiers are Linear Discriminant Analysis (LDA), Sparse Representation Classifiers (SRC), and Support Vector Machines (SVM). LDA-based methods seek linear combinations of features to best separate the two groups. Wolzet et al. performed various training using MR images that combined multiple features such as hippocampal volume, tensor-based shape measures, and cortical thickness. Cho et al. classified AD patients using an incremental method utilizing cerebral cortical thickness data with spatial frequency components to which the manifold harmonic transform was applied.

[0007] In contrast, SRC-based methods have attempted to find sparse linear combinations with minimized representation residuals. Xu et al. developed a multimodal classification framework that includes three modalities: volumetric MRI, fluorodeoxyglucose, and positron emission tomography (PET). Chenet et al. proposed a group discriminative sparse representation algorithm that analyzes the efficiency of group label information.

[0008] Finally, in the SVM-based method, nonlinear decision boundaries were used to maximize the separation of MCI-C and MCI-N using the kernel trick. Zhang et al. proposed a 3D discrete wavelet transform-based MRI preprocessing method and verified its validity for the early diagnosis of AD and MCI. Weiet et al. distinguished MCI-C and MCI-N using a combination of MRI features derived from FreeSurfer and node features derived from a thickness network.

[0009] Recently, various methods applying deep learning have been developed, and in particular, prediction results based on CNNs (Convolutional Neural Networks) are demonstrating high accuracy. The convolutional layer, which is the core of a CNN, has the advantage of sharing locally connected parameters. This helps the network correctly assume image characteristics (statistical stationarity and locality of pixel dependency). LeNet, AlexNet, VGGNet, and GoogLeNet are considered representative CNN architectures due to their success in the ImageNet competition. However, these architectures suffer from a significant problem called vanishing gradients during deep model training. ResNet is more widely used than the aforementioned CNN architectures by solving this problem through shortcut connections, which allow inputs from lower layers to be used in upper layers. Furthermore, as CNN architectures have been improved to learn not only flat images but also stereoscopic images, it has become easier to accurately predict AD transformations in MCI using the brain images themselves.

[0010] Despite these successes, predicting AD transition remains a challenging task due to several reasons related to the variability of the subject groups. First, there is minimal inter-group difference in brain MRI data between the MCI-C and MCI-N groups. As a single category of MCI, both groups are generally viewed as an intermediate state between cognitive normal (CN) and AD. Unlike the classification of MCI versus CN or AD, distinguishing differences in brain volume between MCI-C and MCI-N is a difficult problem. Second, both groups exhibit significant within-group variability. The criteria for MCI are not strict, and the subject range is broad, leading to significant individual variations in symptoms. Furthermore, the correlation between the severity of MCI and the risk of transition to AD is not clear. For these reasons, brain volumes can vary significantly among subjects within the same group, making large differences within the group inevitable. Consequently, these issues make it difficult to find appropriate classifiers to distinguish whether MCI patients will transition to AD. Prior art literature

[65535] ZHAO, Yan, et al. Prediction of Alzheimer's disease progression with multi-information generative adversarial network. IEEE Journal of Biomedical and Health Informatics, 2020, 25.3: pp.711-719.(2020.07.03.) The problem to be solved

[0011] The present invention aims to provide a prospective classification device for predicting dementia and a method of operation thereof, which predicts the risk of a patient with mild cognitive impairment transitioning into a patient with dementia by converting features of diagnostic brain imaging data acquired at the time of diagnosis of a patient with mild cognitive impairment into features of prognostic brain imaging data corresponding to a prognostic time after the time of diagnosis. means of solving the problem

[0012] A prospective classification device for predicting dementia according to an embodiment of the present invention includes at least one processor configured to predict the risk of a patient with mild cognitive impairment transitioning into a patient with dementia by executing a prospective classification program recorded in memory.

[0013] The above-mentioned at least one processor is configured to convert features of diagnostic brain imaging data acquired at the time of diagnosis of the patient with mild cognitive impairment into features of prognostic brain imaging data corresponding to a prognostic time after the time of diagnosis by a prospective classification model; and to predict the risk of the patient with mild cognitive impairment transitioning into the patient with dementia based on the features of the prognostic brain imaging data converted from the features of the diagnostic brain imaging data.

[0014] The above prospective classification model is trained to convert the features of diagnostic brain imaging data acquired for a patient with mild cognitive impairment at a first time point into the features of prognostic brain imaging data acquired for the patient at a second time point after the first time point.

[0015] The above at least one processor may be configured to generate a projection data matrix by transforming a diagnostic brain image data matrix acquired at the time of diagnosis of the patient with mild cognitive impairment using a learned projection matrix of the prospective classification model; generate a prospective data matrix by smoothing the projection data matrix to be adapted to a manifold of a prognostic brain image data matrix using a brain graph matrix of the prospective classification model; and predict the risk of the patient with mild cognitive impairment converting to a dementia patient by applying a learned coefficient vector of the prospective classification model to the prospective data matrix to calculate a dementia conversion risk score representing the probability of mild cognitive impairment converting to dementia.

[0016] The above at least one processor may be configured to generate the brain graph matrix based on a correlation matrix representing connection information of feature regions of the prognostic brain image data matrix and the diagonal matrix of the correlation matrix.

[0017] The above at least one processor generates a projection data matrix by transforming the diagnostic brain image data matrix of each patient with mild cognitive impairment by a projection matrix for training the prospective classification model; produces a brain graph matrix representing a manifold of prognostic brain image data matrices for each of the first patients who have transitioned to dementia after experiencing mild cognitive impairment and the second patients who have not transitioned to dementia after experiencing mild cognitive impairment; generates a prospective data matrix by smoothing the projection data matrix of each of the first patients and the second patients so as to be adapted to the manifold of the prognostic brain image data matrix by the brain graph matrix; generates a divergence function representing the distribution difference between the prospective data matrix generated for each of the first patients and the second patients and the prognostic brain image data matrix of the patient corresponding to each prospective data matrix; and calculates a dementia transition risk score representing the probability that mild cognitive impairment will transition to dementia by a variable including the prospective data matrix and coefficient vector of each of the first patients and the second patients. It may be configured to generate a cross-entropy loss function between the dementia conversion risk score calculated for each of the first patients and the second patients and the dementia conversion correct answer labels of the first patients and the second patients; and to optimize the projection matrix and the coefficient vector based on the derivative generated by partially differentiating the objective function defined by the divergence function and the cross-entropy loss function with respect to the projection matrix and the coefficient vector.

[0018] The above at least one processor may be configured to calculate a first gradient function for the projection matrix of the objective function based on a first derivative of the projection matrix of the cross-entropy loss function and a second derivative of the projection matrix of the divergence function; calculate a second gradient function for the coefficient vector of the objective function based on a third derivative of the coefficient vector of the cross-entropy loss function; and optimize the projection matrix and the coefficient vector by deriving an optimal solution of the projection matrix and the coefficient vector in which the magnitudes of the first gradient function and the second gradient function are simultaneously minimized.

[0019] The at least one processor may be configured to convert the prospective data matrix and the prognostic brain image data matrix into a prospective probability data matrix and a prognostic brain image probability data matrix, respectively, using a softmax function; and to generate a Kullback-Leibler divergence function representing the distribution difference between the prospective probability data matrix and the prognostic brain image probability data matrix.

[0020] The above at least one processor may be configured to generate a first sub-objective function by applying a first coupling coefficient to the cross-entropy loss function; generate a second sub-objective function by applying a second coupling coefficient to the divergence function; generate a normalization item based on the magnitude of the projection matrix and the magnitude of the coefficient vector to reduce the complexity of the projection matrix and the coefficient vector; and generate the objective function based on the first sub-objective function, the second sub-objective function, and the normalization item.

[0021] The first gradient function may include the first derivative, the second derivative, and a linear function of the projection matrix. The second gradient function may include the third derivative and a linear function of the coefficient vector.

[0022] A method of operation of a prospective classification device for predicting dementia according to an embodiment of the present invention includes the step of predicting the risk of a patient with mild cognitive impairment transitioning into a patient with dementia by executing a prospective classification program recorded in memory by at least one processor.

[0023] The above-mentioned predicting step includes: a step of converting the features of diagnostic brain imaging data acquired at the time of diagnosis of the patient with mild cognitive impairment into features of prognostic brain imaging data corresponding to a prognostic time after the time of diagnosis by a prospective classification model; and a step of predicting the risk of the patient with mild cognitive impairment transitioning into the patient with dementia based on the features of the prognostic brain imaging data converted from the features of the diagnostic brain imaging data.

[0024] The above-mentioned transforming step may include: a step of generating a projected data matrix by transforming a diagnostic brain image data matrix acquired at the time of diagnosis of the patient with mild cognitive impairment using a projected matrix learned by the prospective classification model; a step of generating a prospective data matrix by smoothing the projected data matrix to be adapted to the manifold of the prognostic brain image data matrix using a brain graph matrix of the prospective classification model; and a step of predicting the risk of the patient with mild cognitive impairment converting to a dementia patient by applying a learned coefficient vector of the prospective classification model to the prospective data matrix to calculate a dementia conversion risk score representing the probability of mild cognitive impairment converting to dementia.

[0025] The above-mentioned transforming step may further include a correlation matrix representing connection information of feature regions of the above-mentioned prognostic brain image data matrix and a step of generating the above-mentioned brain graph matrix based on the diagonal matrix of the correlation matrix.

[0026] A method of operation of a prospective classification device for predicting dementia according to an embodiment of the present invention may further include the step of learning the prospective classification model based on the diagnostic brain imaging data and the prognostic brain imaging data by the at least one processor.

[0027] The above learning step comprises: generating a projection data matrix by transforming the diagnostic brain image data matrix of each patient with mild cognitive impairment by a projection matrix for the learning of the above prospective classification model; calculating a brain graph matrix representing the manifold of the prognostic brain image data matrix of each first patient who transitioned to dementia after experiencing mild cognitive impairment and the second patient who did not transition to dementia after experiencing mild cognitive impairment; generating a prospective data matrix by smoothing the projection data matrix of each of the first patients and the second patients so as to be adapted to the manifold of the prognostic brain image data matrix by the brain graph matrix; generating a divergence function representing the distribution difference between the prospective data matrix generated for each of the first patients and the second patients and the prognostic brain image data matrix of the patient corresponding to each prospective data matrix; and calculating a dementia transition risk score representing the probability of mild cognitive impairment transitioning to dementia by a variable including the prospective data matrix and coefficient vector of each of the first patients and the second patients. The method may include: a step of generating a cross-entropy loss function between the dementia conversion risk score calculated for each of the first patients and the second patients and the dementia conversion correct answer labels of the first patients and the second patients; and a step of optimizing the projection matrix and the coefficient vector based on derivatives generated by partially differentiating the objective function defined by the divergence function and the cross-entropy loss function with respect to the projection matrix and the coefficient vector.

[0028] The optimization step may include: a step of calculating a first gradient function for the projection matrix of the objective function based on a first derivative of the projection matrix of the cross-entropy loss function and a second derivative of the projection matrix of the divergence function; a step of calculating a second gradient function for the coefficient vector of the objective function based on a third derivative of the coefficient vector of the cross-entropy loss function; and a step of optimizing the projection matrix and the coefficient vector by deriving an optimal solution of the projection matrix and the coefficient vector in which the magnitudes of the first gradient function and the second gradient function are simultaneously minimized.

[0029] The step of generating the above divergence function may include: a step of converting the prospective data matrix and the prognostic brain image data matrix into a prospective probability data matrix and a prognostic brain image probability data matrix, respectively, using a softmax function; and a step of generating a Kullback-Leibler divergence function representing the distribution difference between the prospective probability data matrix and the prognostic brain image probability data matrix.

[0030] The above learning step may further include the step of generating the objective function based on the cross-entropy loss function and the divergence function.

[0031] The step of generating the objective function may include: generating a first sub-objective function by applying a first joint coefficient to the cross-entropy loss function; generating a second sub-objective function by applying a second joint coefficient to the divergence function; generating a normalization item based on the magnitude of the projection matrix and the magnitude of the coefficient vector to reduce the complexity of the projection matrix and the coefficient vector; and generating the objective function based on the first sub-objective function, the second sub-objective function, and the normalization item.

[0032] In addition, according to an embodiment of the present invention, a computer-readable non-transient recording medium is provided, on which a computer program for executing the operation method of the prospective classification device for predicting dementia is recorded. Effects of the invention

[0033] According to an embodiment of the present invention, a prospective classification device for predicting dementia and a method of operation thereof are provided, which predicts the risk of a patient with mild cognitive impairment transitioning into a patient with dementia by converting features of diagnostic brain imaging data acquired at the time of diagnosis of a patient with mild cognitive impairment into features of prognostic brain imaging data corresponding to a prognostic time after the time of diagnosis.

[0034] According to an embodiment of the present invention, by transforming the features of past diagnostic brain imaging data of longitudinal data collected sequentially according to a time series to match the attributes of the features of brain imaging data at a recent time point corresponding to the prognostic time point, the possibility of a patient with mild cognitive impairment transitioning into a patient with dementia can be accurately predicted by a prospective domain adaptation algorithm. Brief explanation of the drawing

[0035] FIG. 1 is a configuration diagram of a prospective classification device for predicting dementia according to an embodiment of the present invention. FIG. 2 is a flowchart of the operation method of a prospective classification device for dementia prediction according to an embodiment of the present invention. FIG. 3 is a conceptual diagram showing the operation of a prospective classification device for predicting dementia according to an embodiment of the present invention. Figure 4 is a flowchart specifically illustrating step S100 of Figure 2. Figure 5 is a flowchart illustrating the step of generating a divergence function in step S110 of Figure 4. Figure 6 is a flowchart specifically illustrating step S140 of Figure 4. FIG. 7 is a flowchart specifically illustrating steps S200 and S300 of FIG. 2. FIG. 8 is an example diagram of diagnostic data and prognostic data to explain the operation method of a prospective classification device according to an embodiment of the present invention. FIGS. 9 and FIGS. 10 are conceptual diagrams illustrating the process of converting diagnostic brain image data into prognostic brain image data according to an embodiment of the present invention. FIG. 11 is a conceptual diagram of a prospective classification model according to an embodiment of the present invention. Figure 12 is an example of a Hammers atlas with 95 ROIs. FIGS. 13 to 15 are graphs showing prospective classification results according to an embodiment of the present invention. Figure 16 shows the ROC curves of five algorithms including the method according to an embodiment of the present invention. Figures 17 and 18 compare the AD conversion prediction results of general training and prospective training using different algorithms. Figure 19 is the result of comparing the performance of the method according to an embodiment of the present invention with that of a CNN. Specific details for implementing the invention

[0036] The advantages and features of the present invention and the methods for achieving them will become clear by referring to the embodiments described below in detail together with the accompanying drawings. However, the present invention is not limited to the embodiments disclosed below but may be implemented in various different forms. These embodiments are provided merely to ensure that the disclosure of the present invention is complete and to fully inform those skilled in the art of the scope of the invention, and the present invention is defined only by the scope of the claims. Throughout the specification, the same reference numerals refer to the same components.

[0037] In this specification, when a part is described as "comprising" a certain component, this means that, unless specifically stated otherwise, it does not exclude other components but may include additional components. As used in this specification, "module," "unit," and "part" refer to a unit that processes at least one function or operation, and may refer to hardware components such as, for example, software, an FPGA, or one or more processors. In describing embodiments of the present invention, if it is determined that a detailed description of related known functions or known configurations may unnecessarily obscure the essence of the present invention, such detailed description is omitted.

[0038] A prospective classification device for predicting dementia and a method of operation thereof according to an embodiment of the present invention converts features of diagnostic brain imaging data acquired at the time of diagnosis of a patient with mild cognitive impairment into features of prognostic brain imaging data corresponding to a prognostic time point after the time of diagnosis by a prospective classification model, and predicts the risk of a patient with mild cognitive impairment transitioning into a patient with dementia based on the features of prognostic brain imaging data converted from the features of diagnostic brain imaging data. To this end, the prospective classification model is trained to convert features of diagnostic brain imaging data acquired for a patient suffering from mild cognitive impairment at a first time point into features of prognostic brain imaging data acquired for a patient after the first time point at a second time point.

[0039] In the case of patients who transition from mild cognitive impairment to dementia, progression to AD occurs after a certain period due to significant brain atrophy; in contrast, patients who do not transition from mild cognitive impairment to dementia show only subtle changes due to normal aging effects. This indicates a difference in brain volume between the two groups at some point in the future. Classification may be easier if longitudinal transitions regarding how a subject's corpus callosum changes from the current diagnosis to the future prognosis can be identified. That is, current brain images are projected into the future along with transition patterns to classify into groups with significant differences. Accordingly, this invention presents a novel method called prospective classification for predicting the transition from mild cognitive impairment (MCI) to dementia (AD). This invention utilizes brain images projected up to the prognostic point, rather than current images at the time of diagnosis, to classify MCI patients into an AD transition group and a non-transition group.

[0040] A method of operation for a prospective classification device for predicting dementia according to an embodiment of the present invention first learns the pattern of brain transition from diagnosis to prognosis for a group transitioning from MCI to AD and a group not transitioning from MCI to AD, respectively. Then, a classifier is trained using transitioned brain volume features that are more clearly distinguishable, and the risk of AD transition for subjects is predicted. The classifier predicts the risk of AD transition for subjects based on the prognostic characteristics of brain volume, which can lead to better performance because the difference between the two groups in prognostic characteristics is more distinct than at the time of diagnosis.

[0041] FIG. 1 is a configuration diagram of a prospective classification device for predicting dementia according to an embodiment of the present invention. FIG. 2 is a flowchart of the operation method of a prospective classification device for predicting dementia according to an embodiment of the present invention. FIG. 3 is a conceptual diagram showing the operation of a prospective classification device for predicting dementia according to an embodiment of the present invention. Referring to FIG. 1 to FIG. 3, a prospective classification device (100) for predicting dementia according to an embodiment of the present invention includes an artificial intelligence learning unit (110), a feature converter (120), and a dementia conversion risk prediction unit (130). The artificial intelligence learning unit (110), the feature converter (120), and the dementia conversion risk prediction unit (130) can perform their functions by executing a prospective classification algorithm (program) stored in memory by at least one processor.

[0042] A method of operation of a prospective classification device for predicting dementia according to an embodiment of the present invention comprises: a step (S100) of learning a prospective classification model based on diagnostic brain image data and prognostic brain image data by an artificial intelligence learning unit (110); a step (S200) of converting features of diagnostic brain image data obtained at the time of diagnosis of a patient with mild cognitive impairment into features of prognostic brain image data corresponding to a prognostic time after the time of diagnosis by a feature converter (120) using the learned prospective classification model; and a step (S300) of predicting the risk of a patient with mild cognitive impairment converting into a patient with dementia based on features of prognostic brain image data converted from features of diagnostic brain image data by a dementia conversion risk prediction unit (130).

[0043] FIG. 4 is a flowchart specifically illustrating step S100 of FIG. 2. Referring to FIG. 1, FIG. 2 and FIG. 4, the step of training a prospective classification model (S100) comprises: generating a prospective data matrix based on a projection matrix and a brain graph matrix based on a diagnostic brain imaging data matrix for each patient with mild cognitive impairment, and generating a divergence function representing the distribution difference between the prospective data matrix and the patient's prognostic brain imaging data matrix (S110); calculating a dementia conversion risk score representing the probability that mild cognitive impairment will convert to dementia by a variable including the prospective data matrix and coefficient vector for each of the first and second patients (S120); and generating a cross-entropy loss function between the dementia conversion risk score calculated for each of the first and second patients and the dementia conversion correct labels of the first and second patients (S130). It may include a step (S140) of optimizing the projection matrix and coefficient vector based on the derivative of the objective function defined by the cross-entropy loss function and the divergence function with respect to the projection matrix and coefficient vector.

[0044] In step S110, the artificial intelligence learning unit (110) can generate a projected data matrix by transforming the diagnostic brain image data matrix of each patient with mild cognitive impairment into a projection matrix for the training of a prospective classification model, and can produce a brain graph matrix representing the manifold of the prognostic brain image data matrix of each first patient who transitioned to dementia after experiencing mild cognitive impairment and the second patient who did not transition to dementia after experiencing mild cognitive impairment. The artificial intelligence learning unit (110) can generate a prospective data matrix by smoothing the projected data matrix of each first patient and the second patient so that it is adapted to the manifold of the prognostic brain image data matrix by the brain graph matrix. The artificial intelligence learning unit (110) can generate a divergence function representing the distribution difference between the prospective data matrix generated for each first patient and the second patient and the prognostic brain image data matrix of the patient corresponding to each prospective data matrix. At this time, the step of calculating the brain graph matrix may include a correlation matrix representing connection information of feature regions of the prognostic brain image data matrix and a step of generating a brain graph matrix based on the diagonal matrix of the correlation matrix.

[0045] FIG. 5 is a flowchart illustrating the step of generating a divergence function in step S110 of FIG. 4. Referring to FIG. 1, FIG. 2, FIG. 4 and FIG. 5, the artificial intelligence learning unit (110) can perform the step (S111) of converting the prospective data matrix and the prognostic brain image data matrix into a prospective probability data matrix and a prognostic brain image probability data matrix, respectively, using a softmax function; and the step (S142) of generating a Kullback-Leibler divergence function representing the distribution difference between the prospective probability data matrix and the prognostic brain image probability data matrix.

[0046] The artificial intelligence learning unit (110) can perform the steps of: generating a first sub-objective function by applying a first joint coefficient to a cross-entropy loss function to generate an objective function; generating a second sub-objective function by applying a second joint coefficient to a divergence function; generating a normalization item based on the magnitude of the projection matrix and the magnitude of the coefficient vector to reduce the complexity of the projection matrix and the coefficient vector; and generating an objective function based on the first sub-objective function, the second sub-objective function, and the normalization item.

[0047] FIG. 6 is a flowchart specifically illustrating step S140 of FIG. 4. Referring to FIG. 1, FIG. 2, FIG. 4 and FIG. 6, the artificial intelligence learning unit (110) may perform the following steps for optimizing the projection matrix and coefficient vector: a step (S141) of calculating a first gradient function for the projection matrix of the objective function based on a first derivative for the projection matrix of the cross-entropy loss function and a second derivative for the projection matrix of the divergence function; a step (S142) of calculating a second gradient function for the coefficient vector of the objective function based on a third derivative for the coefficient vector of the cross-entropy loss function; and a step (S143) of optimizing the projection matrix and coefficient vector by deriving an optimal solution of the projection matrix and coefficient vector in which the magnitudes of the first gradient function and the second gradient function are simultaneously minimized. In an embodiment, the first gradient function may include the first derivative, the second derivative, and a linear function of the projection matrix. The second gradient function may include a third derivative and a linear function of the coefficient vector.

[0048] FIG. 7 is a flowchart specifically illustrating steps S200 and S300 of FIG. 2. Referring to FIG. 1, FIG. 2 and FIG. 7, the feature converter (120) may perform the steps of: generating a projection data matrix by converting a diagnostic brain image data matrix acquired at the time of diagnosis of a patient with mild cognitive impairment by a learning projection matrix of a prospective classification model (S210); and generating a prospective data matrix by smoothing the projection data matrix by a brain graph matrix of a prospective classification model so as to be adapted to a manifold of a prognostic brain image data matrix (S220).

[0049] More specifically, the feature converter (120) can generate a projected data matrix by converting the diagnostic brain image data matrix obtained at the time of diagnosis of a patient with mild cognitive impairment using the projected matrix learned by a prospective classification model, generate a brain graph matrix based on the correlation matrix representing the connection information of feature regions of the prognostic brain image data matrix and the diagonal matrix of the correlation matrix, and generate a prospective data matrix by smoothing the projected data matrix to be adapted to the manifold of the prognostic brain image data matrix by the brain graph matrix of the prospective classification model.

[0050] The dementia conversion risk prediction unit (130) can perform the step (S310) of predicting the risk of a patient with mild cognitive impairment converting to a dementia patient by applying a learned coefficient vector of a prospective classification model to a prospective data matrix generated by a feature converter (120) to calculate a dementia conversion risk score representing the probability that mild cognitive impairment will convert to dementia. Below, a prospective classification device for predicting dementia and a method of operation thereof according to an embodiment of the present invention will be described in more detail.

[0051] The method of operation of a prospective classification device according to an embodiment of the present invention consists of three stages: preprocessing, prediction, and classification. In the preprocessing stage, brain images are converted into brain volume features using a well-established voxel-based morphological measure (VBM). Subsequently, prediction is performed to learn the transition from diagnosis to prognosis, and classification is performed on the risk of AD transition based on the converted features. In the projection stage, KL divergence loss is calculated, and in the classification stage, cross-entropy loss is measured. Learning for prediction and learning for classification are implemented as an end-to-end procedure that combines the two losses into a single objective function.

[0052] During brain conversion, brain volume features are first transformed through a projection matrix that learns longitudinal changes between diagnosis and prognosis. The transformed features are adjusted to various prognostic features. The risk of AD conversion after brain conversion is calculated using conversion and adaptation functions. While any classifier may be used for this purpose, embodiments of the present invention propose a method using logistic regression to simultaneously acquire information on features that primarily contribute to the problem of distinguishing between AD conversion and non-conversion. The coefficient vector β of the logistic regression provides this information.

[0053] The operation method of a prospective classification device according to an embodiment of the present invention boils down to the problem of estimating a projection matrix P and a coefficient vector β. FIG. 8 is an example diagram of diagnostic data and prognostic data to explain the operation method of a prospective classification device according to an embodiment of the present invention. FIGS. 9 and 10 are conceptual diagrams illustrating the process of converting diagnostic brain image data into prognostic brain image data according to an embodiment of the present invention. The diagnostic data matrix corresponding to diagnosis , the prognosis data matrix corresponding to the prognosis It is represented as. To represent the prospective data matrix, the diagnostic data matrix The brain features are linearly transformed by the projection matrix P and adapted by the prognostic brain graph matrix G.

[0054]

[0055] The projection matrix P is determined to minimize the values ​​of the projected features and the actual values ​​of the prognostic features. In the embodiment, the discrepancy between the actual values ​​of the prognostic features corresponding to the prognostic brain imaging data matrix and the values ​​of the projected features (values ​​of the prospective data matrix) can be measured using Kullback-Leibler (KL) divergence, which measures the difference between the two distributions. Thus, each data matrix Z (prospective data matrix) and (Prognostic brain imaging data matrix) uses the softmax function to form probability matrices (Prospective probability data matrix), It is converted into a (prognostic brain imaging probability data matrix). A divergence function representing the difference in distribution between the two distributions of the prospective probability data matrix and the prognostic brain imaging probability data matrix, such as the Kullback-Leibler (KL) divergence, is calculated as follows:

[0056]

[0057] If the two distributions are similar, the KL-divergence approaches 0; otherwise, the KL-divergence increases. Meanwhile, the projected data generated by the simple transformation It provides only the longitudinal change in diagnosis-prognosis. Adaptation to the prognosis manifold may be performed to make the predicted data more similar to the actual prognostic data. The manifold is usually represented as a graph. In an embodiment of the present invention, the brain graph matrix is the prognostic data matrix (prognostic brain imaging data matrix) Projection data to adapt to the manifold Smooths it.

[0058] Brain graph matrix is projection data It provides connection information between ROI features, and the correlation matrix It consists of. The higher the ROI correlation value between i and j of the correlation matrix, the It increases. The brain graph matrix G is a normalized graph. It is defined as a diagonal matrix am.

[0059] Forward transformation is useful for predicting when datasets from different time points will be provided. Furthermore, given the data labels—that is, in the case of supervised learning—it is possible not only to determine the optimal projection matrix P that reduces the loss, which is the difference between the label and the output, but also to determine the longitudinal difference and manifold adaptation. A logistic classifier can be used to classify the transformation from MCI to AD. The transformed features can be used as inputs for the AD transformation risk as shown in the equation below.

[0060]

[0061] Optimal coefficient vector of logistic regression can be determined through least squares estimation. Since Z in Equation (1) is not fixed, the coefficient vector Least squares estimation cannot be applied. The projection data matrix Z is determined by the projection matrix P, and conversely, the projection matrix P is determined again by the projection data matrix Z. This means that the projection matrix P is and This is because it is determined to reduce the loss between them. In an embodiment of the present invention, a cross-entropy loss defined as shown in the following formula is used.

[0062]

[0063] Projection matrix P and coefficient vector of logistic regression To find it, the above formulas and regularizers are combined to utilize an objective function for forward classification such as the following formula.

[0064]

[0065] Here and It penalizes complexity, and and are combination coefficients ( The objective function is the first coupling coefficient to the cross-entropy loss function. The second joint coefficient to the first sub-objective function and divergence function to which is applied The second sub-objective function to which is applied, and normalization items corresponding to the squared magnitude of the projection matrix and the squared magnitude of the coefficient vector and It can be generated as the sum of.

[0066] The objective function is optimized using the gradient descent method. First, the minimization of the projection matrix P is described. To find the gradient with respect to the projection matrix P, the first and second terms of the objective function are derived. The derivative of the normalizer term is ignored. The derivative of the cross-entropy loss is derived as shown in the following formula.

[0067]

[0068] The KL divergence is differentiated as shown in the following formula.

[0069]

[0070] Consequently, the gradient of P (the first gradient function) is given by the following formula.

[0071]

[0072] The first gradient function may include a first derivative of the cross-entropy loss function with respect to the projection matrix, a second derivative of the divergence function with respect to the projection matrix, and a linear function of the projection matrix P. The initial projection matrix P is It is set as the correlation matrix. Next, Explains the minimization of . Coefficient vectors regarding the classifier The derivative of is given by the following formula.

[0073]

[0074] including a normalizer The gradient of (second gradient function) is as follows:

[0075]

[0076] The second gradient function may include a third derivative of the coefficient vector of the cross-entropy loss function and a first function of the coefficient vector. When a prospective classification model according to an embodiment of the present invention learns preprocessed data, the KL divergence loss and the cross-entropy loss are minimized simultaneously, rather than through an end-to-end procedure. P and The optimization for appears complex because the parameters are intertwined. However, by utilizing the property that the two parameters are sequentially associated, a solution with a structure similar to a multi-layer perceptron can be easily obtained. If a structure like the one shown in Fig. 11 is constructed, the coefficient vectors for the classifier are similar to the weight updates in a general neural network. The derivative of and Using the gradient of P and The value is updated. In this way, the projection matrix and the coefficient vector can be optimized by deriving an optimal solution of the projection matrix and the coefficient vector in which the magnitudes of the first gradient function and the second gradient function are simultaneously minimized.

[0077] The prospective classification model is trained through the aforementioned process. The prospective classification model is trained to convert diagnostic MRI into something similar to prognostic MRI, and then optimized to accurately predict AD conversion for the converted MRI. After model training and optimization, new MCI patient data can be applied to the model to predict AD conversion. In this case, the new data requires only the MRI at the time of diagnosis. The model converts the input new MRI into future data. Then, it calculates the risk of AD conversion for the converted virtual MRI. Therefore, the method of operation of the prospective classification device according to the embodiment of the present invention can diagnose AD conversion in MCI patients early without directly verifying longitudinal changes through follow-up observation.

[0078] FIG. 11 is a conceptual diagram of a prospective classification model according to an embodiment of the present invention. Brain metastasis represents brain volume characteristics expected through linear transformation from diagnosis to prognosis and smoothing of the prognosis by a brain graph. The risk of AD conversion is a coefficient vector based on the transformed features. is estimated and predicted by the logistic classifier. P and Optimization for is the input ( )-Potential( )-Potential( )-output of power( It is solved by constructing a structure similar to an MLP network with 4 layers. , , and The weight matrices are the projection matrix P, brain graph G, and coefficient vector, respectively. It corresponds to. G is determined by the brain graph fixed matrix, not by the learned parameters. Coefficient vector It provides information on key features that contribute to AD conversion / non-conversion risk.

[0079] Hereinafter, an experiment for verifying the performance of a prospective classification device and method according to an embodiment of the present invention is described. The data used in the experiment of the present invention was obtained from the ADNI (Alzheimer's Disease Neuroimaging Initiative) database. Subjects with two or more records in chronological order were classified as follows according to general criteria used to distinguish between AD and MCI subjects: (a) MCI subjects: MMSE score between 24 and 30, memory impairment with memory loss measured by the education-adjusted score on the Wechsler Memory Scale Logical Memory II, Clinical Dementia Rating (CDR) 0.5, (b) AD subjects: MMSE score between 20 and 26, CDR 0.5 or 1.0. As a result, a total of 147 subjects were selected, including 73 MCI-C and 74 MCI-N, and this is summarized in Table 1.

[0080]

[0081] A total of 294 MR images were collected from selected subjects, categorized by diagnosis and prognosis. The MR images were preprocessed to extract ROI-based features. First, AC (anterior commissure)-PC (posterior commissure) correction was performed on all images. Then, the structural MR images were segmented into three different tissues—gray matter (GM), white matter (WM), and cerebrospinal fluid (CSF)—using the Computational Anatomy Toolbox for the SPM package (CAT12: http: / / www.neuro.uni-jena.de / cat / ). For brain images containing the three segmented tissues, subject-labeled images were obtained using the BrainNet Viewer based on the Hammers atlas, which consists of 95 ROIs as shown in Figure 3. Finally, the GM tissue volumes of the 95 ROIs, normalized to the total intracranial volume estimated by summing the GM, WM, and CSF volumes of all ROIs, were used as features for the given subjects.

[0082] Next, the results for unidirectional domain adaptation are described. Figure 12 is an example of a Hammers atlas with 95 ROIs. The method according to an embodiment of the present invention has two objectives: reducing inconsistencies between distributions of brain volume data (distribution matching) and making the relationships between brain ROIs in two domains similar (connection mapping). The results for each objective are presented in order.

[0083] The results of the prospective classification are illustrated in Figures 13 to 15. Figure 13 shows the results of the prospective classification, illustrating the effect of the brain ROI conversion by the projection matrix. The eight axes included in the radial diagram represent the super-categories of the 95 ROIs. Figure 13 demonstrates the effect of the brain conversion in which the 95 ROIs are divided into eight super-categories: frontal lobe (FL), occipital lobe (OL), parietal lobe (PL), cerebellum (CB), ventricle (VT), subcortical lobe (SC), cingulate cortex (CC), and temporal lobe (TL). This change demonstrates that the conversion from diagnosis to prognosis allows for a clearer distinction between the patterns of the two groups in the feature space. This change implies that the conversion from diagnosis to prognosis generates a feature space capable of more clearly distinguishing the patterns of the two groups.

[0084] In Fig. 13, the blue area represents the difference in original functions for each category of MCI-C and MCI-N, and the red area represents the difference in transformed features. Consequently, while they overlap and are difficult to distinguish before transformation, the two groups are clearly distinguishable after transformation. This implies that the transformed ROI has more discriminative power than the ROI at diagnosis. This means that the projection matrix has successfully learned brain ROI feature representations that transform the two groups from diagnosis to prognosis. The method according to an embodiment of the present invention provides more separable results in terms of the distribution of prediction scores for AD conversion risk, as illustrated in Figs. 14 and 15. In an embodiment of the present invention, the AD conversion risk score The distribution of is well differentiated as shown in Figures 14 and 15.

[0085] To compare the accuracy of AD conversion prediction, an experiment was designed to measure the effect of converted features using four competing algorithms: Logistic Regression (LGR), Linear Discriminant Analysis (LDA), Sparse Representation Classifier (SRC), and Support Vector Machine (SVM). First, AD conversion in MCI patients was predicted by directly learning diagnostic data using these algorithms (general learning). Second, predictions were performed by training converted features through the method according to an embodiment of the present invention (prospective training). The entire experiment was conducted by repeating 5-fold cross-validation 100 times, and prediction accuracy was measured by the Area Under Received (AUC). The overall comparison results are shown in Figures 16, 17, and 18. Figure 16 shows the overall AUC comparison results using individual ROC curves. Figure 17 shows the overall performance improvement using converted features. Figure 18 is a scatter plot showing individual results comparing the AUC with and without converted features. Points located on the diagonal indicate that the performance is better than that of the algorithm on the vertical axis.

[0086] Figure 16 shows the ROC curves of five algorithms including the method according to an embodiment of the present invention. In the case of the embodiment of the present invention, the AUC is 0.881, indicating performance 34.7% to 88.7% higher than other algorithms (LGR, LDA, SRC, SVM). Figures 17 and 18 compare the AD conversion prediction results of normal training and prospective training using other algorithms (LGR, LDA, SRC, SVM). Figure 17 provides an overall performance improvement due to the conversion effect. The converted features were fed into the inputs of LGR, LDA, SRC, and SVM. The AUC of each algorithm improved by an average of approximately 26.5%. Figure 18 shows the individual performance improvement according to prospective training in each iteration. Points in the scatter plot located on the diagonal indicate that the algorithms on the vertical axis perform better. Figure 18 shows that most points are located on the diagonal. Among the algorithms used as comparison methods, the three algorithms other than LGR were also used in previous studies to predict AD conversion. Coupe et al. used changes in hippocampal volume and MMSE as criteria for AD conversion and predicted them using LDA, yielding an average AUC performance of 0.657. Xu et al. performed SRC-based prediction using MRI and PET as multimodal data, and the performance for the MRI data was 0.506. Wei et al. applied MRI variables selected through feature selection to SVM and obtained an average performance of 0.742. The AUC of 0.881 obtained by the method according to the embodiment of the present invention surpasses existing studies. Figures 5b and 5c experimentally demonstrate that converted features are useful not only for logistic classifiers but also for other algorithms.

[0087] The method according to an embodiment of the present invention was compared with a deep learning-based CNN. The purpose of this experiment was not only to compare prediction accuracy but also to empirically verify the change in the discriminative power of the method according to the number of training samples. To this end, comparative experiments were conducted on datasets reduced by 50% and 20% compared to the original dataset. Among various CNN models, ResNet was selected as the comparison method to avoid the Vanishing Gradient problem. ResNet was trained on voxel-level 3D image data without preprocessing. The model architecture was designed as a network consisting of a total of 71 layers, including 78 connections, 20 convolutional layers, and fully connected layers. Model parameters were optimized using stochastic gradient descent with a learning rate of 0.001. Model performance was measured by classification accuracy over 100 iterations.

[0088] The results of the comparative analysis with CNN are shown in Fig. 19. The green and blue bars represent the classification accuracy of the ResNet trained on diagnostic and prognostic data, respectively, and the red bar represents the results of the method according to an embodiment of the present invention. When the original dataset was initially trained as shown in Fig. 19(a), the method according to an embodiment of the present invention showed performance that was 44.1% and 27.6% better than the ResNet trained on diagnostic and prognostic data, respectively. The higher accuracy of the method according to an embodiment of the present invention was also observed in datasets reduced by 50% and 20%, as shown in Fig. 19(b) and Fig. 19(c). In particular, the accuracy of the proposed method in the 20% reduced dataset was nearly identical to the accuracy of the ResNet trained on the original prediction dataset. Despite the superiority of ResNet, a sufficient amount of data is required to train a large number of parameters. However, collecting a large amount of data is not easy because collecting actual cases of transition from MCI to AD requires years or decades of follow-up observation. This is a common difficulty encountered in medical domain applications, as exemplified in the present invention. Therefore, a lightweight and accurate model capable of training with a small amount of data is inevitable. Compared to ResNet, it can be seen that the method according to the embodiments of the present invention stores more data and exhibits better performance when data is limited. Considering these points, the method according to the embodiments of the present invention has advantages and implies scalability to similar cases.

[0089] As described above, the method according to an embodiment of the present invention performs prospective classification to predict the transition from MCI to AD. The method according to an embodiment of the present invention performs classification using brain images transformed to a prognostic point in time, rather than the current images at the time of diagnosis. This process includes adaptation to various prognostic features. The risk of transition to AD after brain metastasis is calculated using a logistic classifier that simultaneously provides information on primary brain regions contributing to the problem of distinguishing AD transition / non-transition. Experimental results on the ADNI dataset show that the transition group and the non-transition group are more distinct after the transition. Furthermore, it demonstrates that the transformed features can be used in all classifiers and can improve original performance. This implies the extended use of the transition process.

[0090] Several descriptions regarding the method according to the embodiments of the present invention are as follows. First, to perform prospective classification, a dataset containing at least one pair of different time points must be prepared. In this regard, the application of the method according to the embodiments of the present invention may be considered somewhat limited. However, since patient clinical data is primarily obtained through multiple follow-up examinations after a certain period, it will not be difficult to find such cases, and the method is suitable for the characteristics of the medical field. Second, if prospective classification is extended to include continuous transitions across multiple time points rather than just two, its use may become more general. Third, as can be seen from experiments, the transition process can be beneficial to all classifiers. Therefore, using a more sophisticated classifier offers greater room for improvement. Finally, prospective classification is not limited to image data but can be applied to all types of data.

[0091] At least some of the configurations of the embodiments described above may be implemented as hardware components, software components, and / or a combination of hardware components and software components. For example, the devices, methods, and components described in the embodiments may be implemented using one or more general-purpose computers or special-purpose computers, such as, for example, a processor, a controller, an Arithmetic Logic Unit (ALU), a Digital Signal Processor, a microcomputer, a Field Programmable Gate Array (FPGA), a Programmable Logic Unit (PLU), a microprocessor, or any other device capable of executing and responding to instructions.

[0092] The processing unit may execute an operating system and one or more software applications executed on said operating system. Additionally, the processing unit may access, store, manipulate, process, and generate data in response to the execution of the software. For convenience of understanding, the processing unit may be described as being used as a single unit, but a person of ordinary skill in the art will understand that the processing unit may include multiple processing elements and / or multiple types of processing elements.

[0093] For example, the processing unit may include multiple processors or one processor and one controller. Additionally, other processing configurations, such as a parallel processor, are also possible. Software may include a computer program, code, instructions, or a combination of one or more of these, and may configure the processing unit to operate as desired or instruct the processing unit independently or collectively.

[0094] Software and / or data may be embodied in any type of machine, component, physical device, virtual equipment, computer storage medium, or device so as to be interpreted by a processing device or to provide instructions or data to a processing device. Software may be distributed over networked computer systems and stored or executed in a distributed manner. Software and data may be stored on one or more computer-readable recording media.

[0095] The method according to the embodiment may be implemented in the form of program instructions that can be executed through various computer means and recorded on a computer-readable medium. The computer-readable medium may include program instructions, data files, data structures, etc., either alone or in combination. The program instructions recorded on the medium may be those specifically designed and configured for the embodiment, or they may be those known and available to those skilled in the art of computer software.

[0096] Examples of computer-readable recording media include magnetic media such as hard disks, floppy disks, and magnetic tapes; optical recording media such as CD-ROMs and DVDs; and hardware devices specifically configured to store and execute program instructions, such as ROM, RAM, and flash memory. Examples of program instructions include machine code, such as that generated by a compiler, as well as high-level language code that can be executed by a computer using an interpreter, etc. The hardware devices described above may be configured to operate as one or more software modules to perform the operation of the embodiments, and vice versa.

[0097] Although the embodiments have been described above with reference to limited examples and drawings, those skilled in the art can make various modifications and variations from the description above. For example, suitable results may be achieved even if the described techniques are performed in a different order than described, and / or if the components of the described system, structure, device, circuit, etc. are combined or assembled in a form different from described, or replaced or substituted by other components or equivalents. Therefore, other implementations, other embodiments, and equivalents to the claims are also included within the scope of the claims set forth below.

Claims

Claim 1 The apparatus includes at least one processor configured to predict the risk of a patient with mild cognitive impairment transitioning into a dementia patient by executing a prospective classification program recorded in memory, wherein the at least one processor converts features of diagnostic brain imaging data acquired at the time of diagnosis of the patient with mild cognitive impairment into features of prognostic brain imaging data corresponding to a prognostic time after the time of diagnosis by a prospective classification model, and predicts the risk of the patient with mild cognitive impairment transitioning into a dementia patient based on the features of prognostic brain imaging data converted from the features of the diagnostic brain imaging data, wherein the prospective classification model is a model trained to convert features of diagnostic brain imaging data acquired for a patient suffering from mild cognitive impairment at a first time point into features of prognostic brain imaging data acquired for the patient after the first time point at a second time point, and wherein the at least one processor performs preprocessing to convert diagnostic brain imaging data acquired at the time of diagnosis of the patient with mild cognitive impairment into brain volume features using a voxel-based morphological measurement method (VBM), and extracts the tissue volume of each ROI based on a brain atlas composed of a plurality of ROIs (Regions of Interest, anatomical regions of interest). The method is configured to generate a diagnostic brain imaging data matrix, generate a projection data matrix by transforming the diagnostic brain imaging data matrix using a learned projection matrix of the prospective classification model, generate a prospective data matrix by smoothing the projection data matrix using a brain graph matrix of the prospective classification model so as to be adapted to the manifold of the prognostic brain imaging data matrix, and generate a prospective data matrix by applying a learned coefficient vector of the prospective classification model to the prospective data matrix to calculate a dementia conversion risk score representing the probability that mild cognitive impairment will convert to dementia, thereby predicting the risk of a patient with mild cognitive impairment converting to a dementia patient.A prospective classification device for dementia prediction, characterized in that the brain graph matrix is ​​generated based on a correlation matrix representing connection information of feature regions of the prognostic brain image data matrix and the diagonal matrix of the correlation matrix. Claim 2 delete Claim 3 delete Claim 4 In claim 1, the brain graph matrix represents a manifold of prognostic brain imaging data matrices for each of the first patients who transitioned to dementia after experiencing mild cognitive impairment and the second patients who did not transition to dementia after experiencing mild cognitive impairment; the prospective data matrix is ​​generated by smoothing the projection data matrix for each of the first patients and the second patients so as to be adapted to the manifold of the prognostic brain imaging data matrix by the brain graph matrix; the dementia transition risk score represents the probability that mild cognitive impairment will transition to dementia by a variable including the prospective data matrix and coefficient vector for each of the first patients and the second patients; and the at least one processor generates a divergence function representing the distribution difference between the prospective data matrix generated for each of the first patients and the second patients and the prognostic brain imaging data matrix of the patient corresponding to each prospective data matrix, generates a cross-entropy loss function between the dementia transition risk score calculated for each of the first patients and the second patients and the dementia transition correct labels of the first patients and the second patients, and A prospective classification device for dementia prediction configured to optimize the projection matrix and the coefficient vector based on derivatives generated by partially differentiating the objective function defined by the divergence function and the cross-entropy loss function with respect to the projection matrix and the coefficient vector. Claim 5 A prospective classification device for dementia prediction according to claim 4, wherein at least one processor calculates a first gradient function for the projection matrix of the objective function based on a first derivative of the cross-entropy loss function for the projection matrix and a second derivative of the divergence function for the projection matrix; calculates a second gradient function for the coefficient vector of the objective function based on a third derivative of the cross-entropy loss function for the coefficient vector; and optimizes the projection matrix and the coefficient vector by deriving an optimal solution of the projection matrix and the coefficient vector in which the magnitudes of the first gradient function and the second gradient function are simultaneously minimized. Claim 6 A prospective classification device for dementia prediction according to claim 5, wherein the first gradient function comprises the first derivative, the second derivative and a linear function of the projection matrix, and the second gradient function comprises the third derivative and a linear function of the coefficient vector. Claim 7 A prospective classification device for dementia prediction according to claim 4, wherein at least one processor is configured to convert the prospective data matrix and the prognostic brain image data matrix into a prospective probability data matrix and a prognostic brain image probability data matrix, respectively, using a softmax function; and to generate a Kullback-Leibler divergence function representing the distribution difference between the prospective probability data matrix and the prognostic brain image probability data matrix. Claim 8 A prospective classification device for dementia prediction according to claim 4, wherein at least one processor is configured to generate a first sub-objective function by applying a first coupling coefficient to the cross-entropy loss function; generate a second sub-objective function by applying a second coupling coefficient to the divergence function; generate a normalization item based on the magnitude of the projection matrix and the magnitude of the coefficient vector to reduce the complexity of the projection matrix and the coefficient vector; and generate an objective function based on the first sub-objective function, the second sub-objective function, and the normalization item. Claim 9 A method of operation of a prospective classification device for predicting dementia, comprising the step of predicting the risk of a patient with mild cognitive impairment transitioning into a dementia patient by executing a prospective classification program recorded in memory by at least one processor, wherein the predicting step comprises: a step of converting features of diagnostic brain imaging data acquired at the time of diagnosis of the patient with mild cognitive impairment into features of prognostic brain imaging data corresponding to a prognostic time after the time of diagnosis by a prospective classification model; and a step of predicting the risk of the patient with mild cognitive impairment transitioning into a dementia patient based on the features of prognostic brain imaging data converted from the features of the diagnostic brain imaging data; wherein the prospective classification model is a model trained to convert features of diagnostic brain imaging data acquired for a patient suffering from mild cognitive impairment at a first time point into features of prognostic brain imaging data acquired for the patient at a second time point after the first time point; and the converting step comprises a step of performing preprocessing to convert diagnostic brain imaging data acquired at the time of diagnosis of the patient with mild cognitive impairment into brain volume features using a voxel-based morphological measurement method (VBM). A step of generating a diagnostic brain imaging data matrix by extracting the tissue volume of each ROI based on a brain atlas composed of multiple ROIs (Regions of Interest, anatomical regions of interest); a step of generating a projection data matrix by transforming the diagnostic brain imaging data matrix using a learned projection matrix of the prospective classification model; a step of generating a prospective data matrix by smoothing the projection data matrix to be adapted to the manifold of the prognostic brain imaging data matrix using a brain graph matrix of the prospective classification model; and a step of predicting the risk of a patient with mild cognitive impairment converting to a dementia patient by applying a learned coefficient vector of the prospective classification model to the prospective data matrix to calculate a dementia conversion risk score representing the probability of mild cognitive impairment converting to dementia.A method of operation of a prospective classification device for dementia prediction, comprising: a step of generating a brain graph matrix based on a correlation matrix representing connection information of feature regions of the prognostic brain image data matrix and a diagonal matrix of the correlation matrix; and further comprising the transforming step. Claim 10 delete Claim 11 delete Claim 12 In claim 9, the method further comprises the step of learning the prospective classification model based on the diagnostic brain imaging data and the prognostic brain imaging data by the at least one processor; wherein the learning step comprises: generating a projection data matrix by transforming the diagnostic brain imaging data matrix of each patient with mild cognitive impairment by a projection matrix for learning the prospective classification model; calculating a brain graph matrix representing the manifold of the prognostic brain imaging data matrix of each first patient who transitioned to dementia after experiencing mild cognitive impairment and second patient who did not transition to dementia after experiencing mild cognitive impairment; generating a prospective data matrix by smoothing the projection data matrix of each of the first patients and the second patients so as to adapt it to the manifold of the prognostic brain imaging data matrix by the brain graph matrix; generating a divergence function representing the distribution difference between the prospective data matrix generated for each of the first patients and the second patients and the prognostic brain imaging data matrix of the patient corresponding to each prospective data matrix; and the prospective data of each of the first patients and the second patients A method of operation of a prospective classification device for dementia prediction, comprising: a step of calculating a dementia conversion risk score representing the probability that mild cognitive impairment will convert to dementia by means of a variable including a matrix and a coefficient vector; a step of generating a cross-entropy loss function between the dementia conversion risk score calculated for each of the first patients and the second patients and the dementia conversion correct label of the first patients and the second patients; and a step of optimizing the projection matrix and the coefficient vector based on a derivative generated by partially differentiating the objective function defined by the divergence function and the cross-entropy loss function with respect to the projection matrix and the coefficient vector. Claim 13 A method of operation of a prospective classification device for dementia prediction according to claim 12, wherein the optimizing step comprises: a step of calculating a first gradient function for the projection matrix of the objective function based on a first derivative of the cross-entropy loss function for the projection matrix and a second derivative of the divergence function for the projection matrix; a step of calculating a second gradient function for the coefficient vector of the objective function based on a third derivative of the cross-entropy loss function for the coefficient vector; and a step of optimizing the projection matrix and the coefficient vector by deriving an optimal solution of the projection matrix and the coefficient vector in which the magnitudes of the first gradient function and the second gradient function are simultaneously minimized. Claim 14 A method of operation of a prospective classification device for dementia prediction, wherein, in claim 13, the first gradient function comprises the first derivative, the second derivative and a linear function of the projection matrix, and the second gradient function comprises the third derivative and a linear function of the coefficient vector. Claim 15 A method of operation of a prospective classification device for dementia prediction according to claim 12, wherein the step of generating the divergence function comprises: a step of converting the prospective data matrix and the prognostic brain image data matrix into a prospective probability data matrix and a prognostic brain image probability data matrix, respectively, using a softmax function; and a step of generating a Kullback-Leibler divergence function representing the distribution difference between the prospective probability data matrix and the prognostic brain image probability data matrix. Claim 16 A method of operation of a prospective classification device for dementia prediction, wherein the learning step further comprises: a step of generating the objective function based on the cross-entropy loss function and the divergence function; and the step of generating the objective function comprises: a step of generating a first sub-objective function by applying a first coupling coefficient to the cross-entropy loss function; a step of generating a second sub-objective function by applying a second coupling coefficient to the divergence function; a step of generating a normalization item based on the magnitude of the projection matrix and the magnitude of the coefficient vector to reduce the complexity of the projection matrix and the coefficient vector; and a step of generating the objective function based on the first sub-objective function, the second sub-objective function, and the normalization item. Claim 17 A computer-readable, non-transient recording medium having a computer program recorded thereon for executing the method of operation of a prospective classification device for dementia prediction described in claim 9.

Images