A method implemented by computer means for depicting features of at least one observation of interest
Patent Information
- Application Number
- JP2024575292
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2022-06-22
- Filing Date
- 2023-06-20
- Publication Date
- 2025-07-17
AI Technical Summary
Existing deep learning techniques for depicting medical image characteristics require large labeled training datasets and are prone to overfitting, making them inefficient and unreliable for anomaly detection in new images.
A method implemented by computer means that determines a template of normal objects through diffeomorphic deformation and minimizes a cost function to depict characteristics without annotations or large training sets, allowing for efficient anomaly detection.
This method enables effective anomaly detection in medical images without the need for large labeled datasets or annotations, reducing overfitting and improving generalizability, thus providing a robust and efficient solution for characterizing image characteristics.
Smart Images
Figure 2023247573000001 
Figure 2023247573000002 
Figure 2023247573000003
Abstract
Description
Technical Field
[0001] The present invention relates to a method implemented by computer means for depicting the characteristics of at least one observation of an object. The above observation may be an image, for example, a medical image, or a feature of an image. The above object may be a patient, or a part of a patient, for example, an organ of the above patient.
Background Art
[0002] In order to depict the characteristics of medical images, for example, to detect abnormalities such as lesions or tumors in the above images, it is known to apply deep learning techniques.
[0003] Such techniques conventionally involve training a model from training data including an image and an annotation indicating whether an abnormality is seen in each image and what abnormality is seen. This trained model is then used during inference to depict the characteristics of new images that are not present in the training set.
[0004] The main drawbacks of these techniques are that they require a large number of images that make up the training set, and the abnormalities have to be labeled, i.e., have to be labeled by a doctor for each image. Such techniques may also be prone to overfitting, which means that the results obtained with the training model fit too closely to the training data and cannot be generalized on new data, i.e., cannot reliably predict abnormalities on new images.
Prior Art Documents
Non-Patent Documents
[0005]
Non-Patent Document 1
[0006] The present invention proposes a method that requires neither annotations or labels nor a large training set and can be easily generalized to many application examples. [Means for Solving the Problems]
[0007] For this purpose, the present invention proposes a method implemented by computer means for depicting the characteristics of at least one observation y of an object, the method comprising: - determining a template depicting the characteristics of a population of normal objects by diffeomorphic deformation;
[0008] [Number]
[0009] and - minimizing the following cost function J, namely
[0010] [Number]
[0011] wherein provided that F is a deformation function, A is an anomaly matrix, ||A||1 is the 1-norm of A, K and λ are constant values, Reg is a regularization function, wherein, during the minimization, F and A are determined by learning, F provides information regarding the morphological variations of the object, and A provides information regarding the anomalies in the above-mentioned observation y.
[0012] The above-mentioned observation may be an image. The above-mentioned image may be a 2D or 3D image.
[0013] The above-mentioned observation may be a feature of an image, for example, a contour. The above-mentioned contour may be an open or closed contour.
[0014] The above image may be a medical image, for example, an MRI image of a subject, in this case a part of a patient. The above part may be an organ of the above patient.
[0015] The above image may also be a non-medical image. For example, the above image may show the temporal variation of a given feature (such as the temperature of food cooking).
[0016] Template
[0017]
Number
[0018] may be obtained by Large Diffeomorphic Deformation Metric Mapping, i.e., LDDMM.
[0019] The LDDMM framework is known from the following papers. - Alain Trouve. An infinite dimensional group approach for physics based models in pattern recognition. Preprint, 1995. - Paul Dupuis, Ulf Grenander, and Michael I Miller. Variational problems on flows of diffeomorphisms for image matching. Quarterly of applied mathematics, pages 587 - 600, 1998. - M Faisal Beg, Michael I Miller, Alain Trouve, and Laurent Younes. Computing large deformation metric mappings via geodesic flows of diffeomorphisms. International journal of computer vision, 61(2):139 - 157, 2005.
[0020] The construction of templates using the LDDMM framework is also known from the paper "Construction of Bayesian deformable models via stochastic approximation algorithm: A convergence study. S. Allassonniere, E. Kuhn, A. Trouve, Bernoulli Journal, Vol 16(3), 641 - 678, 2010".
[0021] Template
[0022]
Number
[0023] is based on a plurality of observations without anomalies, called control observations, which may be, for example, based on images.
[0024] Using the LDDMM framework, such templates based on multiple images
[0025]
Number
[0026] The determination method will be described in detail below.
[0027] Given a dataset of images in dimension d ∈ {2, 3} (y i ) 1≦i≦n assuming the above, the objective of the method is to create a template
[0028]
Number
[0029] , that is, to create a representative image of the target population of individuals.
[0030] To do so, a distance between observations or images using diffeomorphic deformation is created.
[0031] Let V be a Reproducible Kernel Hilbert Space.
[0032]
Number
[0033] In the case of, the vector field v is represented as follows.
[0034]
Number
[0035] However,[[]]
[0036]
Number
[0037] is called a control point,
[0038]
Number
[0039] is called momentum. Thus, v is, for the core K V is expressed as the interpolation of the momentum at the control point using. In fact, K V is the variance
[0040]
Number
[0041] is chosen to be a Gaussian kernel with, that is,
[0042]
Number
[0043] in,
[0044]
Number
[0045] is.
[0046]
Number
[0047] Let it be.
[0048]
Number
[0049] Assuming,
[0050]
Number
[0051] is set to the diffeomorphism obtained as the flow of the vector field v at time 1, that is,
[0052]
Number
[0053] is.
[0054] Next,
[0055]
Number
[0056] is set as the group of such diffeomorphisms.
[0057] Next, it is possible to define a distance for G. When φ, φ' ∈ G,
[0058]
Number
[0059] and
[0060]
Number
[0061] is set.
[0062] This shows that G is given a manifold structure and the distance on it is calculated as the length of the minimal geodesic path connecting two elements
[0063]
Number
[0064] of.
[0065] This infimum is the minimum value, and it has been shown that the distance is right invariant. Furthermore, the geodesic of G passing through the identity at the initial time is then uniquely defined by the initial velocity v0, i.e., by the initial control point and momentum.
[0066] Below, we write the value of this geodesic at time t as εxp t (v0).
[0067] Thus, for two shapes x and y ∈ M, we set the following.
[0068]
Number
[0069] This distance measures the shortest length of the path from x to y using the diffeomorphism φ v . This distance also makes it possible to define a Riemannian structure for M. The geodesics for M are then defined using the initial shape p0 and the initial velocity v0,
[0070]
Number
[0071] as defined by.
[0072] Thus, the distance between two shapes is measured as a measure of the difficulty of deforming one into the other. Furthermore,
[0073]
Number
[0074] is a diffeomorphism, so it is invertible and preserves the smoothness and structure of the shapes.
[0075] Use their geodesics to define a template for the dataset and the deformations from this template to each object using inaccurate matching. More precisely, set the following cost function.
[0076]
Number
[0077] However, v i is obtained using the control point c0 and the momentum α i .
[0078]
Number
[0079] Thus, it is a template of the population and can be estimated, for example, by minimizing the above cost function using the normal gradient descent algorithm. Other algorithms may also be used.
[0080] Such a template or control template based on multiple anomaly-free images
[0081]
Number
[0082] Once obtained, the following steps or models may be applied.
[0083] This model aims to highlight the anomalies of the object with respect to the above control template. When the anomalies are modeled, it is assumed that these anomalies are scattered in the corresponding parts of the object.
[0084] Except for these abnormalities, the remainder of this part may be the same as the control template. Therefore, the following model is proposed.
[0085] The observation of n objects is (y i ) 1≦i≦n , and the control template obtained using the method described above is
[0086]
Number
[0087] is written as. All images are the same size
[0088]
Number
[0089] and is considered to have, for a 2D or 3D image, d = 2 or 3.
[0090] Each observation y i can be written as follows.
[0091]
Number
[0092] This equation means that the observation y i of each object is the template i with the matrix A i added, including the above-mentioned observation abnormalities and noise ε
[0093]
Number
[0094] obtained as a diffeomorphic deformation of.
[0095] As described above, v iuses the Gaussian kernel as follows
[0096]
Number
[0097] control points
[0098]
Number
[0099] in terms of momentum
[0100]
Number
[0101] is the velocity field obtained as an interpolation of momentum.
[0102]
Number
[0103] :
[0104]
Number
[0105] To emphasize the expected sparsity of anomalies, an L i regularization of A 1 is added. This is also a way to make it noise-free and is modeled as the following centered normal distribution.
[0106] Assuming this model, the goal is to jointly estimate the deformation and anomaly matrices from a given template for each observation. Here, since the control template has already been estimated, each observation can be processed separately, and the following function is minimized.
[0107]
Number
[0108] However, v i is obtained using Equation (E3). The first term of J i measures the distance between the observation y i and the reconstruction
[0109]
Number
[0110] and the other terms
[0111]
Number
[0112] and
[0113]
Number
[0114] measure the sparsity of A i and the regularity of the diffeomorphic deformation, respectively.
[0115] A i To minimize ||.||1, since it is not differentiable, for example, a proximal gradient descent algorithm may be implemented using the Pytorch package to automatically calculate the gradient of the differentiable part of J i .
[0116] Alternatively, the template
[0117]
Number
[0118] may be based on only one observation without anomalies, called control observation, and at least one observation with anomalies.
[0119] In fact, in some cases, it may be difficult to obtain a dataset of control observations necessary to create the control template
[0120]
Number
[0121] For example, this may be the case for the brain where it is hardly possible to obtain MRI scans from control patients.
[0122] In that case, the template
[0123]
Number
[0124] may be created according to the estimation of the anomaly matrix using only the image of one control patient y0. To do so,
[0125]
Number
[0126] is considered to be the diffeomorphic deformation of the control hypertemplate y0.
[0127]
Number
[0128] However, as described above, v0 is obtained as an interpolation of the momentum α0 at the control point c0.
[0129] Since y0 is a control observation, there is no abnormality, and its diffeomorphic deformation
[0130]
Number
[0131] is also normal.
[0132] Therefore, in that case, not only the velocity (v i ) 1≦i≦n and the sparse matrix (A i ) 1≦i≦n are estimated, but also the velocity v0 is estimated by minimization.
[0133]
Number
[0134] However,
[0135]
Number
[0136] is obtained using the velocity v0, and v i is obtained as an interpolation of the momentum α i at the control point c0.
[0137] Since the template is the reference image for the entire population, the energy to be minimized is not separable and includes the contributions of all objects.
[0138] Again, this optimization may be performed by proximal gradient descent.
[0139] To reduce the computation time, the same control point c0 may be used for all velocity fields (v i ) 1≦i≦n . In an alternative form, the control point c i may vary with the velocity field (v i ). 1≦i≦n
[0140] In practice, computational problems may occur when estimating the parameters of those models. In fact, gradient descent often tends to get stuck at local minima. By including reconstruction errors in the singular matrix, the algorithm often chooses not to improve the reconstruction and remains stuck at local minima. To solve this problem, the singular matrix may be made to take on other values of object reconstruction for the first iteration, for example, for the first 100 iterations. This makes it possible to improve the diffeomorphic reconstruction and reach a more relevant region of interest in the energy landscape.
[0141] Below, the above-mentioned model estimates the template
[0142]
Number
[0143] and deformation without using either the singular matrix or the hyper-template and can be compared with a normal cross-sectional atlas. This can be written as the minimization of the following energy.
[0144]
Number
[0145] The above model can also be compared with the cross-sectional atlas when the template is retrieved by the control hyper-template. As described above, v0 is obtained as an interpolation of the momentum α0 at the control point c0. The function to be minimized is as follows here.
[0146] [Number]
[0147] However,
[0148] [Number]
[0149] is defined by Equation E5.
[0150] The estimation of these two models can be performed using ordinary gradient descent.
[0151] The present invention also proposes a computer software, which, when executed by a processor, includes instructions for implementing at least a part of the above method.
[0152] The present invention also proposes a computer device, which - an input interface for receiving at least one observation y of the object, and - a memory for storing at least the instructions of the above computer program, and - a processor for reading the aforementioned instructions and then accessing the memory to execute the above method, and - an output interface for providing F and A determined during the minimization of the cost function J and includes.
[0153] The present invention also proposes a computer-readable non-transitory recording medium on which computer software is registered, and when the computer software is executed by a processor, the above-described method is implemented.
[0154] Other features, details, and advantages are shown in the following embodiments for carrying out the invention and in the figures.
Brief Description of the Drawings
[0155]
Figure 1
Figure 2
Figure 3
Figure 4
Figure 5
Figure 6
Figure 7
Figure 8
Figure 9
Figure 10
Figure 11
Figure 12
Embodiments for Carrying Out the Invention
[0156] Figure 1 schematically shows an example of a computer device 1 according to the present invention. The above computer device 1 includes - an input interface 2, and - a memory 3 for storing at least instructions of a computer program, and - a processor 4 that reads the aforementioned instructions and accesses the memory 3 to execute the method or algorithm according to the present invention, and - an output interface 5 and is provided with.
[0157] Next, the method or algorithm implemented according to the present invention will be described through several examples.
Example
[0158] To test the above model, a simulated dataset of 500 observations deformed from a common template is created, to which a random number of dark spots (between 1 and 5) and some Gaussian noise are added. This template is created as a deformation of an ellipse (hyper-template). Control template
[0159]
Number
[0160] is estimated, and a "control" dataset of 100 observations of interest without dark spots is also created from the same template so that it can be used in the case of the model (E4).
[0161] Five observations with the template and dark spots are presented in the first row of Figure 2.
[0162] More specifically, this figure shows the following. - In the top row, the template and five observations, - In the second row, the estimated template and the reconstructed object (model (E7)) that do not use either the hyper-template or the abnormal matrix. - In the third row, the result when using the hyper-template and not using the abnormal matrix (model (E0)). - In the fourth row, first estimate the template (first column) from the control object, and then reconstruct other objects using the abnormal matrix (model (E4)). - In the last row, the result using the sparse matrix when the template is estimated simultaneously from the hyper-template (model (E6)).
[0163] First apply the model (E7) where the template is directly estimated without using either the hyper-template or the sparse matrix.
[0164] As can be seen in the second row of Figure 2, a dark shadow is created on the estimated template. Similarly, this dark shadow is reported in the reconstruction of each observation. Furthermore, as can be seen in the last two columns, in order to minimize J0, the algorithm may inaccurately estimate the object deformed to exclude the dark spot.
[0165] In the third row of Figure 2, apply the model (E8) where the template is obtained from the elliptical hyper-template but without the abnormal matrix. At this time, as expected, there is no dark shadow on the template, but the reconstruction of the last two observations is still poor.
[0166] The model (E4) is also tested. To do this, the template from 100 "control" observations is estimated. This template is then fixed by minimizing equation (E4). This estimated template and some reconstructions are shown in the fourth row of Figure 2.
[0167] At this time, the observations are better reconstructed and the dark spots are removed. Their intensities are a little weaker than before for the proximal gradient descent that applies a soft threshold to the residuals. Furthermore, the shapes of the dark spots are well identified, and their positions are also well identified.
[0168] Next, Model E6 is applied and the template is estimated from the hyper-template. The results are presented in the last row of Figure 2. Again, the observations are well reconstructed and the dark spots are removed in the anomaly matrix. The shapes, positions, and volumes of the dark spots are captured.
[0169] Finally, the four models described above are compared by calculating the average registration error when considering simply the form of Table 1 (i.e., the error between the observations without dark spots and the reconstructions without the anomaly matrix). As expected, the error is smaller when considering models (E4) and (E6).
[0170]
Table 1
[0171] Therefore, it was possible to reconstruct the observations and extract their anomalies using models (E4) and (E6). Furthermore, the improvement in reconstruction was emphasized when estimating both the anomaly matrix and the deformation.
[0172] The choice to estimate both the deformation v and the anomaly matrix A simultaneously, rather than sequentially, is brought about by the effort to improve the reconstruction of the object. Indeed, in Table 1, it can be seen that the reconstruction error is smaller when estimating both simultaneously. Specifically, it can be seen in the third row of Figure 2, and in the last two columns, the residuals include the entire part of the observation. Those parts are therefore removed in the anomaly matrix.
Example
[0173] The purpose of this example is to emphasize this need to simultaneously estimate both deformation and the anomaly matrix. Different images of that example can be seen in Figure 3.
[0174] At this time, black lines are added to the control template, mimicking, for example, the veins of the liver or the turns of a brain slice. From this template, one observation with dark spots added to it is created. Try to reconstruct this observation from the template with or without the anomaly matrix.
[0175] Figure 3 shows, from left to right, the fixed template, the observation, the reconstruction without the anomaly matrix, and the reconstruction with the anomaly matrix.
[0176] Without the anomaly matrix, the model chooses to severely deform the black lines to create dark spots. Specifically, the part of the line is here in the residual, and the black dots are not completely within it. The estimation of the anomaly matrix from this residual is thus poor. However, when both deformation and the anomaly matrix are estimated simultaneously, the black lines are well registered from the template to the individual, and only the black dots are taken out in the anomaly matrix.
[0177] These two observations support the need to combine the estimation of the anomaly matrix and the deformation.
Example
[0178] In this example, the model is applied to a dataset of brains with tumors obtained from the BraTS 2018 dataset (Bakas et al., 2017, 2018, Menze et al., 2014). More precisely, the dataset consists of 50 post-contrast T1-weighted MRI scans of glioblastoma and low-grade glioma.
[0179] Do not discard the control dataset, but simply discard one control observation. Thus, the model (E6) is used to estimate the template using this control as a hyper-template.
[0180] The purpose is to reconstruct each target brain from the control template and to obtain tumors in the abnormal matrix. Ideally, diffeomorphic deformation registers the brain folds (called gyri) and ventricles. What cannot be removed in diffeomorphic deformation is thus only the tumor.
[0181] Figure 4 shows the template estimated by the above algorithm.
[0182] Figure 5 shows the results for observations of four different targets. For each target, the figure shows, on the left, the observations corresponding to three different planes and, on the right, the estimated abnormal matrix.
[0183] As can be seen, the lesions are extracted in the abnormal matrix with just a small error in the reconstruction. Specifically, the gyri are well registered and do not exist in the abnormal matrix. Regarding the ventricles, if a part of one is present in the lower right image, they are also well registered in the other images. In fact, for the patient in the lower right, his right ventricle is completely different from the template, which poses difficulties for the algorithm to register a part of it. However, in all cases, the use of the LDDMM framework has made it possible to register most of the non-abnormal brain and thus to obtain a clean abnormal matrix.
[0184] Here, an important choice of the parameters in equation (E6) is actually the sparsity constant λ. It can be chosen to take a smaller λ. This makes it possible to include the edema around the tumor (the dark area around the tumor) in the abnormal matrix. However, it may also include more reconstruction errors. Here, λ is chosen to visually detect the tumor with as little reconstruction error as possible even if the whole tumor is not in the abnormal matrix. Such a choice is sufficient to inform the physician of the suspected abnormalities, which are actually included in the abnormal matrix.
[0185] In fact, 59 out of 62 tumors in the dataset are visible in the anomaly matrix (95%). For the tumors that are not visible, the tumors are small lesions in the highly variable zones of the brain. In such cases, as shown in Figure 6, the tumor (the enclosed area) is not easy to distinguish and is in the middle of the gyrus.
Example
[0186] In this example, the method does not require a large dataset to be applied and does not require annotations for the location of the tumor. To emphasize this advantage, the exact same algorithm is applied to only one brain with a tumor. For the template, it is fixed as a brain without a tumor. Therefore, only two different brains are used to try to detect anomalies. Specifically, compare the anomaly matrix and results estimated for this patient in the previous section before the templates were estimated together (see Figure 7).
[0187] Figure 7 shows the observations on the left, the anomaly matrix estimated with the template fixed as the control target in the center, and the anomaly matrix when estimating the template on the right, as done in Example 3. The results are shown in two different slices, namely, the location of the tumor (the upper image) and the upper part of the brain (the lower image). In both cases, the tumor is removed. When not estimating the template, more reconstruction errors are included.
[0188] The tumor is also removed in the anomaly matrix with small reconstruction errors, specifically at the boundary and upper part of the brain. Specifically, the error at the upper part does not exist when estimating the template together with the anomaly matrix. In fact, the variability between the control target and the target with a tumor is greater there, resulting in larger reconstruction errors. However, estimating the template prevents this problem even in the case of a small number of subjects as done in Example 3.
[0189] Even if a sparse abnormal matrix contains many errors, it is emphasized that since it was generated using only two objects and since the method according to the invention can produce valid results without access to more objects.
Example
[0190] In this example, the algorithm of the invention is applied to a dataset of the liver. The goal is to reconstruct the liver of each patient while restoring the tumor with an abnormal matrix. Since there is no access to the complete dataset of the control patients, model (E4) cannot be used and it is necessary to estimate the template from the hyper-template using model (E6). The hyper-template is selected as a patient or object without a tumor, not from the database.
[0191] Two different structures, namely the tumor as a dark spot and the blood vessels as white structures, are visible in the liver. Therefore, it is those two structures that are restored with the abnormal matrix. If one simply wants to remove the tumor, it is easy to separate the tumor from the blood vessels according to the intensity of the tumor. Furthermore, the noise level can be completely different for each object. If the dataset has not been preprocessed, an efficient sparsity constant λ cannot be found for each object, and in those with the highest level of noise, this noise is restored with the abnormal matrix. To prevent this phenomenon, it is first decided to convolve the observations with a Gaussian kernel. The resulting image is shown in Figure 8.
[0192] Figure 9 shows, on the left, the first image or observation and, on the right, the same image after convolution using the above Gaussian kernel.
[0193] This smoothing makes it possible to have a robust algorithm for this population. Furthermore, since not all images have the same pixel interval, some parts of the image may be downsampled. Also, all of the images may be included in a black box of the same size.
[0194] The results for different subjects are presented below.
[0195] As a post - processing step, the coefficients of the anomaly matrix are set to 0 except for the reconstruction and the target. If this selection is not made, reconstruction errors are reported in the anomaly matrix. Specifically, white zones of the anomaly matrix in voxels where the diffeomorphic deformation could not reproduce the liver part, and black zones where the diffeomorphic deformation created a liver part in a place where there should be no liver part will be found.
[0196] The template estimated as the diffeomorphic deformation of the control patient
[0197]
Number
[0198] is presented in Figure 10. This estimation benefits in particular from the use of the entire dataset of the control patient. Thus, since it is derived from a specific control subject, the vessels of this specific subject still exist in the final template and may affect future estimations of the anomaly matrix. Having a template of the control subject also ensures a reduction of the reconstruction error and enables better detection of anomalies.
[0199] Next, Figure 11 shows the results for a subject with no vessels visible by the scanner.
[0200] Specifically, in this figure, the observations are shown on the left (three different planes), the reconstruction is shown in the center, and the anomaly matrix is shown on the right.
[0201] The algorithm according to the present invention can extract tumors in the anomaly matrix. The registered contour also exists in the anomaly matrix as it is used to obtain a better final reconstruction.
[0202] Figure 12 shows a subject with visible vessels. From left to right, this figure shows the observation, the anomaly matrix, and the negative values of the anomaly matrix.
[0203] As can be seen, not only the tumor but also the blood vessels are removed. If one wants to find only the tumor, the first thing that comes to mind is to examine the negative values of the anomaly matrix. To extract only the anomalies without small errors in the reconstruction, further investigation regarding post-processing is required.
[0204] Finally, Figure 13 shows the importance of applying Gaussian convolution to the dataset before estimating the model parameters.
[0205] More specifically, Figure 12 shows - In the top row, the results when the dataset is not pre-convolved with a Gaussian kernel, - Below, the above results with pre-processing, - On the left, the observations, - In the center, the reconstructions, - On the right, the anomaly matrix are shown.
[0206] If this pre-processing is not performed, noise may be extracted in the anomaly matrix and the tumor may not be visible. This problem is actually solved after convolution, and the tumor (in the upper left of the liver) is removed.
[0207] To measure the quality of detection, an MD radiologist segmented the tumors of 10 patients. This led to the segmentation of a total of 133 tumors. Here, since the tumors are dark spots on the liver, only the negative coefficients of the sparse matrix are examined to measure the quality of detection.
[0208] Choose not to evaluate the segmentation other than detection. In fact, here, since it is rare for the algorithm to segment the entire tumor, the dice score is the average. However, here, the goal is not to accurately segment the tumor but simply to inform the physician of the possible anomalies and, in particular, to detect very small lesions.
[0209] When 133 tumors are segmented, the algorithm according to the present invention detects 125 (94%) of them. For the tumors that are not detected, there are two possibilities. Sometimes, the difference between the tumor intensity and the noise is actually small, and the algorithm cannot separate them for some objects where the noise is still high in detail. The other possibility is when the diffeomorphic registration of the liver is not complete and the tumor is outside of it. In that case, the tumor is actually in the abnormal matrix but is lost due to the reconstruction error.
[0210] Finally, not only tumors are taken out in the abnormal matrix. As shown above, small reconstruction errors may exist at the boundary. However, the algorithm also plays its role by detecting other abnormalities besides lesions. Specifically, in some objects, some slightly dark spots are taken out, which are actually due to perfusion disorders.
[0211] The above example shows that the residual of the diffeomorphic deformation from the control template can be used to detect and segment organ lesions. Furthermore, the above example seems to be able to even improve the diffeomorphic reconstruction of observations. This method has the advantage of not requiring a large dataset of diseased patients nor annotations from doctors. Therefore, it is particularly suitable for anomaly detection for specific treatment protocols where it is often impossible to obtain a large dataset in detail. The effectiveness of this method is also shown in a dataset of brains with gliomas and a dataset of livers. Specifically, in the former case, the method according to the present invention was able to register the gyri and ventricles of the brain while extracting the tumor in the abnormal matrix.
Explanation of symbols
[0212] 1 Computer device 2 Input interface 3 Memory 4 Processor 5 Output interface
Claims
1. A method implemented by computer means for depicting the characteristics of at least one observation y of an object, comprising: - determining a template for depicting the characteristics of a population of objects without anomalies by means of diffeomorphic deformation; 【Number 1】 - minimizing the following cost function J, namely where 【Number 2】 wherein F is a deformation function, A is an anomaly matrix, K and λ are predefined constants, Reg is a regularization function, ||A|| 1 is the 1-norm of A, and during the minimization, F and A are determined together, F provides information regarding the morphological variations of the object, and A provides information regarding the anomalies in the observation y.
2. The method according to claim 1, wherein the observation is an image.
3. The method according to claim 1, wherein the observation is a feature of an image, such as a contour.
4. The method according to claim 2 or 3, wherein the image is a medical image.
5. The method according to claim 1, wherein the template is obtained using a large diffeomorphic deformation mapping.
6. 【Number 3】 The method according to claim 1, wherein the template is based on a plurality of observations without anomalies.
7. 【Number 4】 The method according to claim 1, wherein the template is based on only one observation without anomalies and at least one observation with anomalies.
8. 【Number 5】 Computer software comprising instructions that, when executed by a processor, perform at least a part of the method according to claim 1.
9. A computer device (1), comprising: - an input interface (2) for receiving at least one observation y of an object; - a memory (3) for storing at least the instructions of the computer program according to claim 8; - a processor (4) for reading the instructions and then accessing the memory (3) to execute the method according to claim 1; - an output interface (5) for providing F and A determined during the minimization of the cost function J.
10. A computer-readable non-transitory recording medium on which the computer software is registered for implementing the method according to claim 1 when the computer software is executed by a processor (4).