Medical image processing device and medical image processing method

Abstract

To accurately detect a plurality of feature points in a medical image.SOLUTION: According to an embodiment, a medical image processing device includes a reference point detection part, a candidate point generation part, a position score part, a structure score part, and an output part. The reference point detection part detects a reference point having a spatial correlation with a target point in a medical image for each of a plurality of target points corresponding to a plurality of feature points. The candidate point generation part generates a plurality of candidate points corresponding to a target point for each of the plurality of target points by using a detection model. The position score part selects a candidate point for each of the plurality of target points on the basis of a position feature showing a spatial position relation between the target point and the reference point. The structure score part selects a candidate point combination for each of a plurality of candidate point combinations on the basis of a structure feature showing a spatial structure relation among the plurality of target points. The output part outputs a plurality of feature points in the medical image on the basis of the selected candidate point and candidate point combination.SELECTED DRAWING: Figure 1

Description

[Technical Field]

[0001] Embodiments disclosed herein and in the drawings relate to a medical image processing apparatus and a medical image processing method. [Background technology]

[0002] When performing image processing on medical images of a subject, feature point detection is crucial. In particular, when performing image processing such as alignment, segmentation, and quantitative analysis on medical images, accurate detection of feature points in the medical image is required, and other image processing is performed on the medical image based on the detected feature points.

[0003] Conventional techniques involve generating a learning model using machine learning, including deep learning, and then predicting the location of feature points in medical images based on that model. However, such methods have the problem of difficulty in detecting feature points in specific medical images and a low feature point detection rate.

[0004] For example, in conventional technology, one method for detecting accurate target feature points in medical images involves using a stable, clearly defined, and easily detectable target, such as a bone, as a reference point. This is done by eliminating abnormally positioned candidate feature points from a group of detected candidate feature points based on spatial constraints, such as distance limitations between the target feature point and the reference point. Another example of conventional technology involves using machine learning methods to generate a combined detection model capable of predicting multiple target feature points. In the learning phase, the combined detection model learns the features of the target feature points themselves, as well as the relative relationships between multiple target feature points, thereby simultaneously predicting the positions of multiple target feature points.

[0005] However, in methods that detect feature points based on reference points, the definition of spatial constraints is not achieved with high precision. For example, if there is a lot of interfering tissue near a target feature point, such as the intestines being located near the end of the pancreas, multiple similar target feature points that all satisfy the predefined spatial constraints may be detected simultaneously. As a result, the detected point may not be the desired target feature point, leading to a low feature point detection rate. Furthermore, methods that predict target feature points based on the relative relationships between multiple target feature points can only be applied when the positional relationships between the multiple target feature points are stable. For example, human organs such as soft tissues exhibit individual differences, so their positions vary greatly in medical images of different subjects. Therefore, in such cases, using the relative relationships between multiple target feature points does not guarantee the accurate detection of each feature point, resulting in a low feature point detection rate. [Prior art documents] [Patent Documents]

[0006] [Patent Document 1] U.S. Patent No. 10297027 [Patent Document 2] U.S. Patent No. 10210613 [Patent Document 3] Japanese Patent Publication No. 2009-129237 [Patent Document 4] Japanese Patent Publication No. 2016-36685 [Patent Document 5] Japanese Patent Publication No. 2015-36123 [Non-patent literature]

[0007] [Non-Patent Document 1] "Collaborative Regression-based Anatomical Landmark Detection," [Retrieved April 20, 2022], Internet<URL:https: / / www.ncbi.nlm.nih.gov / pmc / articles / PMC4833677 / > [Non-Patent Document 2] "Robust Multi-scale Anatomical Landmark Detection in Incomplete 3D-CT Data," [Retrieved April 20, 1992], Internet <URL:https: / / www5.informatik.uni-erlangen.de / Forschung / Publikationen / 2017 / Ghesu17-RMA.pdf> [Overview of the project] [Problems that the invention aims to solve]

[0008] One of the problems that the embodiments disclosed herein and in the drawings aim to solve is the accurate detection of multiple feature points in medical images. However, the problems that the embodiments disclosed herein and in the drawings aim to solve are not limited to the above problem. Problems corresponding to the effects of each configuration shown in the embodiments described later can also be positioned as other problems. [Means for solving the problem]

[0009] The medical image processing apparatus according to the embodiment is a medical image processing apparatus for detecting a plurality of feature points in a medical image, and comprises a reference point detection unit, a candidate point generation unit, a position score unit, a structure score unit, and an output unit. The reference point detection unit detects a reference point in the medical image that has a spatial correlation with the target point for each of the plurality of target points corresponding to the plurality of feature points. The candidate point generation unit generates a plurality of candidate points corresponding to each of the plurality of target points using a detection model. The position score unit scores a position score for each of the plurality of target points based on position features that indicate the spatial positional relationship between the target point and the reference point in the medical image, and selects a candidate point whose position score satisfies a preset condition. The structure score unit scores a structure score for each of the plurality of candidate point combinations obtained by combining the candidate points among the plurality of target points based on structure features that indicate the spatial structural relationship between the plurality of target points in the medical image, and selects a candidate point combination whose structure score satisfies a preset condition. The output unit outputs the plurality of feature points based on the candidate points selected by the position score unit and the combination of candidate points selected by the structure score unit. [Brief explanation of the drawing]

[0010] [Figure 1] Figure 1 is a block diagram showing the functional configuration of a medical image processing apparatus according to the first embodiment. [Figure 2] Figure 2 is a flowchart showing the steps of the feature distribution calculation stage according to the first embodiment. [Figure 3] Figure 3 is a flowchart showing the calculation of the location feature distribution according to the first embodiment. [Figure 4] Figure 4 is a schematic diagram showing an example of a spatial positional relationship that follows a specific distribution according to the first embodiment. [Figure 5] Figure 5 is a flowchart showing the calculation of the structural feature distribution according to the first embodiment. [Figure 6]Figure 6 is a flowchart showing the steps of the detection stage according to the first embodiment. [Figure 7] Figure 7 is a flowchart showing the steps of the detection stage according to the first embodiment. [Figure 8] Figure 8 is a block diagram showing the functional configuration of a medical image processing apparatus according to the second embodiment. [Figure 9] Figure 9 is a flowchart showing the steps of the classifier learning phase according to the second embodiment. [Figure 10] Figure 10 is a flowchart showing the steps of the detection stage according to the second embodiment. [Figure 11] Figure 11 is a flowchart showing the steps of the detection stage according to the third embodiment. [Figure 12] Figure 12 is a schematic diagram showing an example of a location feature distribution according to the first embodiment. [Modes for carrying out the invention]

[0011] The embodiments of the medical image processing apparatus and medical image processing method will be described in detail below with reference to the drawings.

[0012] A medical image processing apparatus according to one embodiment is a medical image processing apparatus for detecting a plurality of feature points in a medical image, comprising a reference point detection unit, a candidate point generation unit, a position score unit, a structure score unit, and an output unit. The reference point detection unit detects a reference point in the medical image that has a spatial correlation with the target point for each of the plurality of target points corresponding to the plurality of feature points. The candidate point generation unit generates a plurality of candidate points corresponding to each of the plurality of target points using a detection model. The position score unit scores a position score for each of the plurality of target points based on position features that indicate the spatial positional relationship between the target point and the reference point in the medical image, and selects a candidate point whose position score satisfies a preset condition. The structure score unit scores a structure score for each of the plurality of candidate point combinations obtained by combining the candidate points among the plurality of target points based on structure features that indicate the spatial structural relationship between the plurality of target points in the medical image, and selects a candidate point combination whose structure score satisfies a preset condition. The output unit outputs the plurality of feature points based on the candidate points selected by the position score unit and the combination of candidate points selected by the structure score unit.

[0013] According to a medical image processing apparatus according to one embodiment, the position score unit selects a plurality of candidate points whose position scores satisfy the conditions from a plurality of candidate points generated by the candidate point generation unit. The structure score unit selects a candidate point combination whose structure score satisfies the conditions from a plurality of candidate point combinations obtained by combining the plurality of candidate points selected by the position score unit. The output unit outputs the plurality of candidate points included in the candidate point combination selected by the structure score unit as the plurality of feature points.

[0014] According to a medical image processing apparatus according to one embodiment, the structure score unit selects a candidate point combination whose structure score satisfies the conditions from a plurality of candidate point combinations obtained by combining a plurality of candidate points generated by the candidate point generation unit. The position score unit selects a plurality of candidate points whose position score satisfies the conditions from a plurality of candidate points included in the candidate point combination selected by the structure score unit. The output unit outputs the plurality of candidate points selected by the position score unit as the plurality of feature points.

[0015] This allows for the simultaneous and accurate detection of multiple feature points in medical images by assigning positional and structural scores to multiple candidate points of a target point based on their positional and structural features, and selecting candidate points whose positional and structural scores meet predetermined conditions as feature points.

[0016] A medical image processing apparatus according to one embodiment further comprises a feature distribution calculation unit that calculates a location feature distribution showing the probability distribution of the location features and a structure feature distribution showing the probability distribution of the structure features based on the training dataset of the detection model. The location score unit scores the location score for the candidate points based on the location features and the location feature distribution corresponding to the location features. The structure score unit scores the structure score for the candidate point combination based on the structure features and the structure feature distribution corresponding to the structure features.

[0017] A medical image processing apparatus according to one embodiment further comprises a feature distribution calculation unit that calculates a combined position feature distribution showing the probability distribution of combined position features obtained by combining a plurality of spatial positional relationships as position features, and a combined structure feature distribution showing the probability distribution of combined structure features obtained by combining a plurality of structural features as structure features, based on the training dataset of the detection model. The position score unit scores the position score for the candidate points based on the combined position features and the combined position feature distribution corresponding to the combined position features. The structure score unit scores the structure score for the candidate point combination based on the combined structure features and the combined structure feature distribution corresponding to the combined structure features.

[0018] This allows for the calculation of positional and structural features based on positional feature distributions and structural feature distributions, or combined positional and combined structural feature distributions, enabling more accurate selection of candidate points as feature points, and further improving the detection rate of multiple feature points in medical images.

[0019] A medical image processing apparatus according to one embodiment further comprises a classifier learning unit that learns a position feature classifier based on the position features and the spatial positional relationship between anomalies in the medical image and the reference points, and learns a structural feature classifier based on the structural features, the spatial structural relationship between a plurality of anomalies in the medical image, and the spatial structural relationship between a plurality of target points and a plurality of anomalies in the medical image. The position score unit scores a position score for the candidate points based on the position feature classifier. The structure score unit scores a structure score for the candidate point combinations based on the structure feature classifier.

[0020] This allows for further improvement in the detection rate of multiple feature points in medical images by assigning scores and structural scores to multiple candidate points of the target point based on a location feature classifier and a structural feature classifier. In particular, even in medical images where constructing the feature distribution of the target point is difficult, multiple feature points in the medical image can be accurately detected by selecting accurate candidate points as feature points using the location feature classifier and the structural feature classifier.

[0021] According to a medical image processing apparatus of one embodiment, the candidate point generation unit further generates visual characteristics for each of the plurality of candidate points for each of the plurality of target points using the detection model. The position scoring unit scores an individual overall score for each of the candidate points as the position score based on the position characteristics and the visual characteristics. The structure scoring unit scores a combined overall score for each of the candidate point combinations as the structure score based on the structure characteristics and the individual overall scores.

[0022] This allows for a comprehensive evaluation of multiple candidate points for the target point based on visual, positional, and structural features, thereby further improving the detection rate of multiple feature points in medical images. In particular, even when the positional score of interference points in the medical image of a subject is higher than the positional score of the target point due to causes such as surgery, accurate candidate points can be selected as feature points using a comprehensive score based on visual features, thereby enabling accurate detection of multiple feature points in medical images.

[0023] A medical image processing device according to one embodiment further includes an update unit that updates the positional feature distribution and the structural feature distribution by online learning based on the positional score and the structural score.

[0024] A medical image processing device according to one embodiment further includes an update unit that updates the position feature classifier and the structure feature classifier by online learning based on the position score and the structure score.

[0025] This allows for online training of the detection model using the update unit, and by updating the detection model, it is possible to obtain positional and structural scores that are better adapted to the current dataset, thereby further improving the detection rate of multiple feature points in medical images.

[0026] A medical image processing method according to one embodiment is a medical image processing method for detecting a plurality of feature points in a medical image, and includes a reference point detection step, a candidate point generation step, a position score step, a structure score step, and an output step. The reference point detection step detects a reference point in the medical image that has a spatial correlation with the target point for each of the plurality of target points corresponding to the plurality of feature points. The candidate point generation step generates a plurality of candidate points corresponding to each of the plurality of target points using a detection model. The position score step scores a position score for each of the plurality of target points based on position features that indicate the spatial positional relationship between the target point and the reference point in the medical image, and selects a candidate point whose position score satisfies a preset condition. The structure score step scores a structure score for each of the plurality of candidate point combinations obtained by combining the candidate points among the plurality of target points based on structure features that indicate the spatial structural relationship between the plurality of target points in the medical image, and selects a candidate point combination whose structure score satisfies a preset condition. The output step outputs the plurality of feature points based on the candidate points selected in the position score step and the candidate point combination selected in the structure score step.

[0027] According to a medical image processing method according to one embodiment, the position score step selects a plurality of candidate points whose position score satisfies the conditions from a plurality of candidate points generated in the candidate point generation step. The structure score step selects a combination of candidate points whose structure score satisfies the conditions from a plurality of candidate points selected in the position score step. The output step outputs the plurality of candidate points included in the candidate point combination selected in the structure score step as the plurality of feature points.

[0028] According to a medical image processing method according to one embodiment, the structure score step selects a combination of candidate points from a plurality of candidate points generated in the candidate point generation step whose structure score satisfies the conditions. The position score step selects a plurality of candidate points from a plurality of candidate points included in the candidate point combination selected in the structure score step whose position score satisfies the conditions. The output step outputs the plurality of candidate points selected in the position score step as the plurality of feature points.

[0029] A medical image processing method according to one embodiment further includes a feature distribution calculation step of calculating a location feature distribution showing the probability distribution of the location features and a structure feature distribution showing the probability distribution of the structure features based on the training dataset of the detection model. The location score step scores the location score for the candidate points based on the location features and the location feature distribution corresponding to the location features. The structure score step scores the structure score for the candidate point combination based on the structure features and the structure feature distribution corresponding to the structure features.

[0030] This allows for accurate detection of multiple feature points in medical images by assigning positional and structural scores to multiple candidate points of a target point based on their positional and structural features, and by accurately selecting candidate points that meet the required conditions for positional and structural scores.

[0031] According to the medical image processing apparatus and medical image processing method of this embodiment, a position score and a structural score are assigned to multiple candidate points of a target point based on a position feature distribution and a structural feature distribution, and candidate points whose position score and structural score satisfy predetermined conditions are selected as feature points, thereby enabling accurate detection of multiple feature points in a medical image. In particular, a structural score is assigned to a combination of candidate points containing multiple candidate points based on structural features, and candidate points within the combination of candidate points whose structural score satisfies predetermined conditions are selected as feature points, thereby enabling accurate simultaneous detection of multiple feature points in a medical image.

[0032] Furthermore, according to the medical image processing apparatus and medical image processing method of this embodiment, the detection rate of multiple feature points in a medical image can be further improved by scoring positional scores and structural scores for multiple candidate points of a target point based on a positional feature classifier and a structural feature classifier. In particular, even in medical images where it is difficult to construct a feature distribution of the target point, multiple feature points in the medical image can be accurately detected by selecting accurate candidate points as feature points using the positional feature classifier and the structural feature classifier.

[0033] Furthermore, according to the medical image processing apparatus and medical image processing method of this embodiment, the detection rate of multiple feature points in a medical image can be further improved by comprehensively evaluating multiple candidate points of the target point based on visual features, positional features, and structural features. In particular, even when the positional score of interference points in the medical image of a subject is higher than the positional score of the target point due to causes such as surgery, multiple feature points in the medical image can be accurately detected by using a comprehensive score based on visual features to select accurate candidate points as feature points.

[0034] In the medical image processing apparatus and medical image processing method according to this embodiment, a "feature point" is a digitized or encoded point in a medical image that has fixed, essential, and important feature attributes. A "target point" is a point detected as a detection result by the medical image processing apparatus and medical image processing method according to this embodiment for a given feature point. The target point corresponds to the feature point. A "candidate point" is a point generated as a prediction result by the detection model 1 described later for a given target point. One or more candidate points can be generated for a single target point, and one candidate point can be selected as the target point from among the multiple candidate points by evaluating the multiple candidate points based on the position score and structure score described later. A "reference point" is a point whose position is fixed, has a prominent feature, and has a spatial correlation with the target point. The reference point is used in the detection model 1 described later to generate multiple candidate points for the target point as output data. There may be one or more reference points. For example, in the abdominal cavity, the reference point may be the vertebrae or the upper part of the liver, and in the pelvis, the reference point may be the femur or the pubis. The reference point is not limited to these, but can be any point that is fixed in position, has distinctive features, and has a spatial correlation with the target point.

[0035] In the medical image processing apparatus and medical image processing method according to this embodiment, a target point detection model 1 is acquired based on machine learning, and multiple candidate points for the target point are generated. Specifically, in the training stage of the detection model 1, a reference point and a target point are set in the medical image as the training dataset 2 for each target point, and the detection model 1 corresponding to the target point is trained by having it learn the visual features associated with that target point. Furthermore, in the prediction stage, multiple candidate points for the target point are generated by using the trained detection model 1 to detect the target point based on the medical image input as input data. In addition, in the medical image processing apparatus and medical image processing method according to this embodiment, the reference point in the medical image input as input data may be automatically detected, or the reference point in the medical image may be manually specified, and it is sufficient that there is a correlation between the structural position specified as the reference point by an automatic or manual method and the target point.

[0036] (First embodiment) The medical image processing apparatus and medical image processing method according to the first embodiment will be described below with reference to Figures 1 to 7.

[0037] First, the functional configuration of the medical image processing apparatus according to the first embodiment will be described based on Figure 1. Figure 1 is a block diagram showing the functional configuration of the medical image processing apparatus according to the first embodiment.

[0038] As shown in Figure 1, the medical image processing apparatus according to the first embodiment comprises a detection model 1, a reference point detection unit 10, a candidate point generation unit 20, a position score unit 30, a structure score unit 40, a feature distribution calculation unit 50, and an output unit 60. The medical image processing apparatus according to the first embodiment may further include an update unit (details of which will be described later) which is not shown.

[0039] The medical image processing device sets target points and reference points in a medical image, which serves as the training dataset 2, to train the detection model 1. Using the trained detection model 1, it generates multiple candidate points as output data for the medical image input as input data.

[0040] The feature distribution calculation unit 50 calculates a spatial feature distribution based on the spatial positional relationship between target points and reference points in the training dataset 2, and calculates a structural feature distribution based on the spatial structural relationship between multiple target points in the training dataset 2.

[0041] The reference point detection unit 10 detects a reference point that corresponds to the target point.

[0042] The candidate point generation unit 20 generates multiple candidate points for a specified target point based on the trained detection model 1.

[0043] The location scoring unit 30 scores a location score for each of the multiple candidate points based on the location feature distribution and selects candidate points that satisfy the pre-set conditions.

[0044] The structural scoring unit 40 scores a structural score for each of the multiple candidate point combinations based on the structural feature distribution and selects a candidate point combination that satisfies the pre-set conditions.

[0045] The output unit 60 outputs the selected candidate points as multiple feature points in the medical image.

[0046] The medical image processing device then performs various image processing operations on the medical image, such as alignment, segmentation, and quantitative analysis, based on the multiple feature points output by the output unit 60.

[0047] Next, the feature distribution calculation steps according to the first embodiment will be described based on Figures 2 to 5. Figure 2 is a flowchart of steps T100 to T200 of the feature distribution calculation steps according to the first embodiment. Figure 3 is a flowchart of the calculation of the location feature distribution according to the first embodiment. Figure 4 is a schematic diagram showing an example of a spatial positional relationship following a specific distribution according to the first embodiment. Figure 5 is a flowchart of the calculation of the structural feature distribution according to the first embodiment.

[0048] As shown in Figure 2, in step T100, the positional feature distribution of the target point is calculated.

[0049] More specifically, as shown in Figure 3, in step T110, multiple spatial positional relationships between target points and reference points in the training dataset 2 are calculated. Specifically, the feature distribution calculation unit 50 calculates multiple spatial positional relationships between reference points and target points set in the training dataset 2. The process then proceeds to step T120.

[0050] In step T120, a spatial feature is selected from multiple spatial relationships between the target point and the reference point in the training dataset 2. Specifically, the feature distribution calculation unit 50 selects a spatial relationship that follows a specific distribution from the multiple spatial relationships between the target point and the reference point in the training dataset 2 calculated in step T110 as the spatial feature of the target point. Here, we will explain an example where the spatial relationship follows a normal distribution.

[0051] In medical imaging, a certain spatial relationship usually exists between a target point and a reference point due to the anatomical structure of the subject. Here, for example, the space of a medical image is a three-dimensional space of XYZ, where the X, Y, and Z axes are orthogonal to each other. In this case, the positional characteristics of the target point are the spatial relationship between the target point and the reference point, and include the relative displacement of the two in the X, Y, and Z directions, the Euclidean distance between the two in the XY, XZ, and YZ planes, the angle between the two in the XY, XZ, and YZ planes, the Euclidean distance between the two in the three-dimensional space of XYZ, and the result of normalizing these positional characteristics. Normalization is, for example, normalizing the relative displacement of the target point and the reference point in the X, Y, and Z directions with respect to the size of the subject area.

[0052] Using Figure 4 as an example, we will explain how to select spatial relationships that follow a specific distribution as location features. The upper part of Figure 4 shows the spatial relationship of the relative displacement in the Z direction between the target point and the reference point in training dataset 2, and the lower part of Figure 4 shows the spatial relationship of the Euclidean distance in the XY plane between the target point and the reference point in training dataset 2.

[0053] As shown in the upper part of Figure 4, the spatial positional relationship between the target point and the reference point in the Z direction, which is the relative displacement, follows a normal distribution (where both kurtosis and skewness are close to 0). On the other hand, as shown in the lower part of Figure 4, the spatial positional relationship between a similar target point and a reference point in the training data, which is the Euclidean distance in the XY plane, does not follow a normal distribution. Therefore, in this case, the feature distribution calculation unit 50 selects the spatial positional relationship between the target point and the reference point in the Z direction in the training data as a positional feature. Note that one or more positional features may be selected here. For example, if there are multiple spatial positional relationships that follow a particular distribution, the feature distribution calculation unit 50 may select each of the multiple spatial positional relationships that follow the particular distribution as a positional feature. Next, proceed to step T130.

[0054] In step T130, the location feature distribution is calculated. Specifically, the feature distribution calculation unit 50 calculates the probability density function of the distribution followed by the location feature selected in step T120 as the location feature distribution of that location feature. For example, the feature distribution calculation unit 50 uses maximum likelihood estimation to calculate the mean μ and variance σ of the location feature selected in step T120 for the target point and reference point in the training dataset 2. 2 We estimate the position and calculate the probability density function f(x) corresponding to the position feature as the position feature distribution using the following equation (1).

[0055]

number

[0056] Let's illustrate this with an example using Figure 12. Figure 12 is a schematic diagram showing an example of a location feature distribution according to the first embodiment.

[0057] In Figure 12, R1 and R2 represent the reference points R1(x1,y1,z1) and R2(x2,y2,z2) specified in training dataset 2, and A represents the target point A(x) specified in training dataset 2. A ,y A ,z A This shows that there are three spatial positional relationships F1, F2, and F3 between the target point A and the reference points R1 and R2. Here, F1 represents the relative displacement in the Z direction between the target point A and the reference point R1, F2 represents the angle in the XY plane between the target point A and the reference point R2, and F3 represents the Euclidean distance in the XY plane between the target point A and the reference point R2. Also, f1(x), f2(x), and f3(x) represent the probability density functions of the distributions followed by the spatial positional relationships F1, F2, and F3, respectively. Here, the spatial positional relationship F1 is given by F1~N(μ1, σ1 2 Since it follows a normal distribution, the feature distribution calculation unit 50 selects the spatial positional relationship F1 as the positional feature F1 of target point A, and calculates the probability density function f1(x) of the positional feature F1 as the positional feature distribution. Similarly, the feature distribution calculation unit 50 determines whether the spatial positional relationships F2 and F3 follow a specific distribution, and if they do, selects them as the positional features F2 and F3 of target point A, and calculates the corresponding probability density functions f2(x) and f3(x) as the positional feature distribution. Here, the positional feature distribution represents the probability that target point A appears at different positions in the training dataset 2, determined based on the positional features F1, F2, and F3.

[0058] In this way, through steps T110 to T130, the feature distribution calculation unit 50 selects the spatial positional relationship between a target point and a reference point that follows a specific distribution in the training dataset 2 of the detection model 1 as a positional feature, and calculates a probability density function that shows the distribution of said positional feature as the positional feature distribution.

[0059] Here, we have explained an example where the spatial relationship follows a normal distribution. However, you may also select a spatial relationship that follows a different type of distribution as a location feature and calculate the probability density function of that distribution. It is sufficient if the spatial relationship between the target point and the reference point in training dataset 2 follows a specific distribution and that this distribution can be represented by a probability density function.

[0060] Alternatively, by combining multiple spatial positional relationships as positional features, a combined positional feature can be constructed for multiple spatial positional relationships, and the probability density function of this combined positional feature may be calculated as the combined positional feature distribution.

[0061] Returning to Figure 2, in step T200, the structural feature distributions of multiple target points are calculated.

[0062] The relative positions of multiple feature points can reflect the spatial structural relationships between those feature points. For example, the relative positions of multiple feature points within the same organ can reflect the overall structure of the organ, such as its geometric shape and internal morphology. Conversely, the relative positional relationships between multiple feature points in different organs can reflect the anatomical structures of specific parts of the subject's body.

[0063] Based on this, the feature distribution calculation unit 50 selects the spatial structural relationships between multiple target points that follow a specific distribution in the training dataset 2 as structural features, based on the training dataset 2 of the detection model 1, and calculates a probability density function that shows the distribution of said structural features as the structural feature distribution.

[0064] More specifically, as shown in Figure 5, first, in step T210, multiple spatial structural relationships between multiple target points in the training dataset 2 are calculated. Specifically, the feature distribution calculation unit 50 calculates multiple spatial structural relationships between multiple target points in the training dataset 2. Here, the spatial structural relationships between multiple target points are similar to the spatial positional relationships between target points and reference points in step T110, and represent the positional relationships between multiple target points. Next, the process proceeds to step T220.

[0065] In step T220, structural features are selected from multiple spatial structural relationships between multiple target points in the training dataset 2. Specifically, the feature distribution calculation unit 50 selects a spatial structural relationship that follows a specific distribution as a structural feature from the multiple spatial structural relationships between multiple target points in the training dataset 2 calculated in step T210. Here, the spatial structural relationship that follows a specific distribution is similar to that in step T120. The process then proceeds to step T230.

[0066] In step T230, the structural feature distribution is calculated. Specifically, the feature distribution calculation unit 50 calculates the probability density function of the distribution followed by the structural feature selected in step T220 as the structural feature distribution of that structural feature.

[0067] Here, steps T210 to T230 will be explained using the example of a case where the training dataset 2 contains three target points A, B, and C. In this case, in step T210, the feature distribution calculation unit 50 combines target points A, B, and C, and for each pair of target points included in the combination, calculates multiple positional relationships between those two target points as multiple spatial structural relationships. Furthermore, the feature distribution calculation unit 50 selects from the calculated multiple spatial structural relationships the spatial structural relationships that follow a specific distribution as structural features. For example, the feature distribution calculation unit 50 selects structural features F4, F5, and F6. Then, for these structural features F4, F5, and F6, the feature distribution calculation unit 50 calculates the corresponding probability density functions f4(x), f5(x), and f6(x) as structural feature distributions, respectively. Here, the structural feature distributions f4(x), f5(x), and f6(x) represent the probabilities that different structures, composed of multiple combinations of target points A, B, and C determined based on structural features F4, F5, and F6, will appear at different locations in the training dataset 2.

[0068] Alternatively, a combined structural feature may be constructed for multiple spatial structural relationships by combining them as structural features, and the probability density function of this combined structural feature may be calculated as the combined structural feature distribution.

[0069] Thus, in steps T100 to T200, which are feature distribution calculation steps, the feature distribution calculation unit 50 selects location features and structural features based on the target points and reference points in the training dataset 2 of the detection model 1, and calculates the corresponding location feature distribution and structural feature distribution.

[0070] Next, the detection steps according to the first embodiment will be described based on Figures 6 and 7. Figures 6 and 7 are flowcharts showing steps S100 to S600 of the detection steps according to the first embodiment.

[0071] As shown in Figure 6, in step S100, a reference point is detected.

[0072] Specifically, the reference point detection unit 10 detects a reference point in the medical image input as input data that has a spatial correlation with the target point for each of the multiple target points corresponding to the multiple feature points. Here, the reference point detection unit 10 may detect the reference point in the medical image based on the detection model 1, or it may detect or specify the reference point based on other automatic or manual methods. Also, there may be one or more reference points. Furthermore, the one or more reference points that are detected or specified only need to show a correlation with the target point. Next, the process proceeds to step S200.

[0073] In step S200, multiple candidate points for the target point are generated.

[0074] Specifically, the candidate point generation unit 20 uses the detection model 1 to generate a plurality of candidate points corresponding to each of the plurality of target points. Here, the learned detection model 1 generates a plurality of candidate points that most match the visual features of the target point in the medical image input as input data for the specified target point, and outputs them to the position score unit 30. Note that the medical image processing apparatus and the medical image processing method according to the present embodiment simultaneously detect a plurality of feature points, that is, a plurality of target points. Therefore, in step S200, the candidate point generation unit 20 generates a plurality of candidate points for each of the plurality of target points. That is, each of the plurality of target points corresponds to a plurality of candidate points. Subsequently, the process proceeds to step S300.

[0075] In step S300, based on the position feature distribution, a position score is given to the candidate points.

[0076] Specifically, for each of the plurality of target points, the position score unit 30 gives a position score to the candidate points generated by the candidate point generation unit 20 in step S200 based on the position features selected by the feature distribution calculation unit 50 and the calculated position feature distribution in steps T110 to T130 for calculating the position feature distribution.

[0077] Here, regarding the scoring of the position score, the above-mentioned target point A and the position features F1, F2, F3 corresponding to the target point A will be described as an example. For example, in step S200, the candidate point generation unit 20 generates three candidate points D A1 , D A2 , D A3 for the target point A. In that case, the position score unit 30 calculates the position features F1, F2, F3 corresponding to the target point A for the candidate point D A1 (x DA1 , y DA1 , z DA1 ). First, the position score unit 30 calculates, based on the position feature F1, the relative displacement d in the Z direction between the candidate point D A1 (x DA1 , y DA1 , z DA1 ) and the reference point R1(x1, y1, z1)zF1 =z D1 -z1 is calculated. Here, the positional feature F1 is F1~N(μ1, σ1 2 Since it follows a normal distribution, the position score unit 30 uses the position feature distribution corresponding to the position feature F1, i.e., the probability density function f1(x), to calculate the position score S of candidate point DA1 by the following equation (2). DA1_F1 Calculate.

[0078]

number

[0079] Here, f1(dz1) is the candidate point D in the spatial feature distribution. A1 The probability density at is shown, and f1(μ1) represents the probability density at the mean of the training dataset 2 in the location feature distribution. That is, the location score S DA1_F1 Based on the positional feature F1, candidate points D for target point A are located on the curve of the probability density function f1(x). A1 This reflects the distance to the average value of target point A in training dataset 2. Position score S DA1_F1 It is a value greater than 0 and less than or equal to 1, and the position score S DA1_F1 The higher the value (closer to 1), the more candidate points D for target point A are generated based on the positional feature F1. A1 This shows that the distance to the mean value of target point A in training dataset 2 becomes shorter.

[0080] Furthermore, the position score unit 30, based on the position features F2 and F3 corresponding to the target point A, similarly determines candidate point D A1 For this, a location score is calculated based on location features F2 and F3, and the location score S DA1_F2 S DA1_F3 The position score unit 30 calculates the candidate point D using the following equation (3). A1 Total position score S DA1 Calculate.

[0081]

number

[0082] Here, λ represents the weighting coefficient. The position score section 30 is a position score S based on position features F1, F2, and F3. DA1_F1 S DA1_F2 By assigning different weight coefficients λ1, λ2, and λ3 to each point and adding the position scores weighted by each weight coefficient, the candidate point D is obtained. A1 Total position score S DA1 The system calculates the position score. For example, the position score unit 30 assigns higher weight coefficients to position features with high confidence. Here, high confidence means, for example, that the position feature follows a Gaussian distribution or has a small standard deviation.

[0083] Candidate point D A1 Total position score S DA1 The higher the value, the more candidate points D generated by the candidate point generation unit 20 in step S200. A1 This indicates that point A is likely to be the target point. Similarly, the position score unit 30 determines that candidate point D A2 , D A3 A position score will also be assigned to each of them.

[0084] In the above explanation, the position score for a candidate point was calculated by taking the ratio of the probability density at the candidate point in the position feature distribution to the probability density at the mean value of training dataset 2 in the position feature distribution. However, the position score may be calculated using other methods. The position score should be able to indicate the likelihood that a candidate point is the target point based on the position feature distribution. For example, the position score is an index value that indicates the degree of agreement between a candidate point and the target point, or the likelihood that a candidate point is the target point.

[0085] In this way, the position scoring unit 30 assigns a position score to multiple candidate points based on the position feature distribution. Next, the process proceeds to step S400.

[0086] In step S400, a candidate point is selected whose position score meets the pre-set conditions.

[0087] Specifically, the position score unit 30 selects candidate points from a plurality of candidate points generated by the candidate point generation unit 20 in step S200, based on a preset threshold for each of the multiple target points, whose position score exceeds the threshold, and outputs them to the structure score unit 40.

[0088] In step S400, the pre-set condition is that the position score of the candidate point exceeds a pre-set threshold, but it is not limited to this. The position score unit 30 may select several candidate points with the highest position scores from all candidate points and output them to the structure score unit 40. The position score unit 30 may also select candidate points based on the position score using other pre-set criteria, as long as the criteria allow for the selection of candidate points that are likely to become the target point. Next, the process proceeds to step S500.

[0089] In step S500, a structural score is assigned to each candidate point combination based on the structural feature distribution.

[0090] Specifically, the structure score unit 40 scores a structure score for each of the multiple candidate point combinations obtained by combining the candidate points selected by the position score unit 30 in step S400 among multiple target points, based on the structure features selected by the feature distribution calculation unit 50 in steps T210 to T230 and the calculated structure feature distribution.

[0091] Here, regarding the scoring of the structural score, the aforementioned target scores A, B, and C, and candidate scores D corresponding to target scores A, B, and C, respectively. A1 , D A2 , D A3 , D B1 , D B2 , D B3 , D C1 , D C2 , D C3And, structural features F4, F5, and F6 corresponding to target points A, B, and C will be explained as an example. For example, in step S400, the position score unit 30 determines candidate point D for target point A based on the position scores of target points A, B, and C. A1 , D A2 Select candidate point D for target point B. B1 , D B2 Select the candidate point D relative to the target point C. C1 , D C2 Let's say we select this. Here, a candidate point combination means that one combination contains one candidate point corresponding to target points A, B, and C, that is, by combining these candidate points, (D A1 ,D B1 , D C1 ), (D A1 ,D B1 , D C2 ), (D A1 ,D B2 , D C1 ), (D A1 ,D B2 , D C2 ), (D A2 ,D B1 , D C1 ), (D A2 ,D B1 , D C2 ), (D A2 ,D B2 , D C1 ), (D A2 ,D B2 , D C2 ) can be obtained as eight candidate point combinations. In that case, the structural score unit 40 can obtain a certain candidate point combination n(D An ,D Bn , D Cn For a given set of points (n∈[1,8]), structural features F4, F5, and F6 are calculated. Subsequently, the structural score unit 40 calculates the structural score S of candidate point combination n using the structural feature distributions f4(x), f5(x), and f6(x) corresponding to the structural features F4, F5, and F6. n_F4 S n_F5 S n_F6Each is calculated. Then, the structure score unit 40 assigns different weight coefficients to the structure scores based on each structure feature, and adds the structure scores weighted by each weight coefficient to calculate the total structure score S of the candidate point combination n n Here, since the specific calculation method of the structure score is the same as that in step S300, the description is omitted.

[0092] Candidate point combination n (D An , D Bn , D Cn ) The higher the total structure score Sn of the candidate point combination, the higher the possibility that each candidate point D An , D Bn , D Cn among the candidate point combinations is the target point A, B, or C. In the same way, the structure score unit 40 scores the structure scores for other candidate point combinations as well.

[0093] In this way, the structure score unit 40 scores the structure scores for the candidate point combinations based on the structure feature distribution. Subsequently, proceed to step S600.

[0094] In step S600, select the candidate point combinations whose structure scores meet the preset conditions.

[0095] Specifically, the structure score unit 40 selects the candidate points included in the candidate point combination with the highest structure score based on the preset threshold value and outputs them to the output unit 60.

[0096] For example, among the eight candidate point combinations described above, if the structure score of the candidate point combination (D A2 , D B1 , D C2 ) is the highest score, the structure score unit 40 selects the candidate points D A2 , D B1 , D C2 ) included in the candidate point combination (D A2 , D B1 , D C2 respectively.

[0097] In step S600, the pre-set condition is that the candidate point combination has the highest structural score, but it is not limited to this. The structural score unit 40 may select candidate points included in several of the top candidate point combinations with the highest structural scores among all candidate point combinations. The structural score unit 40 may also select candidate point combinations based on the structural score, using other criteria as pre-set conditions, as long as the criteria allow for the selection of candidate points that are likely to be the target points. Next, the process proceeds to step S700.

[0098] In step S700, the selected candidate points are output as feature points.

[0099] Specifically, the output unit 60 outputs the candidate points selected by the structural score unit 40 in step S600 as feature points.

[0100] In the first embodiment, the order of scoring the position score and scoring the structure score may be reversed. That is, as shown in Figure 7, in step S200, the candidate point generation unit 20 outputs the generated candidate points to the structure score unit 40, and then the structure score unit 40 executes step S500 and scores a structure score for multiple candidate point combinations that combine the candidate points generated by the candidate point generation unit 20 based on the structure feature distribution. Next, the structure score unit 40 executes step S600 and selects candidate points included in the candidate point combination with a high structure score from the multiple candidate point combinations and outputs them to the position score unit 30. After that, the position score unit 30 executes step S300 and scores a position score for the candidate points selected by the structure score unit 40 based on the position feature distribution. Next, the position score unit 30 executes step S400 and selects candidate points with a high position score and outputs them to the output unit 60. Then, the output unit 60 outputs the candidate points selected by the position score unit 30 as feature points.

[0101] Furthermore, the more target points there are, the more candidate points there are, and therefore the larger the number of candidate point combinations. To avoid a combinatorial explosion due to such an increase in the number of candidate point combinations, in step S500, the structural score unit 40 may hierarchically score structural scores. Specifically, multiple target points may be divided into several subsets according to the organ or tissue to which they belong. For example, the first subset includes target points corresponding to characteristic points of the liver, the second subset includes target points corresponding to characteristic points of the kidney, and the third subset includes target points corresponding to characteristic points of the pancreas. In these subsets, candidate points corresponding to the target points are combined to calculate their structural scores. At this time, the structural features represent the physiological structural features of each organ or tissue. After scoring and selecting candidate point combinations in different subsets, the structural features between organs or tissues are calculated based on the physiological structural features between organs or tissues, and the selected candidate points are combined to score structural scores and select candidate points.

[0102] According to the medical image processing apparatus and medical image processing method of the first embodiment described above, a positional feature distribution can be calculated based on the spatial positional relationship between a target point and a reference point, and a structural feature distribution can be calculated based on the spatial structural relationship between multiple target points. Furthermore, based on the positional feature distribution and the structural feature distribution, positional scores and structural scores are scored for multiple candidate points of a given target point. Then, candidate points with high positional and structural scores are selected as target points, and these target points are output as feature points. This makes it possible to accurately detect multiple feature points in a medical image.

[0103] Furthermore, according to the medical image processing apparatus and medical image processing method of the first embodiment, since positional scores and structural scores are scored for multiple candidate points corresponding to any one target point based on the positional feature distribution and structural feature distribution, even if similar interference tissues that also follow the same spatial positional relationship exist near the target point, by scoring positional scores and structural scores simultaneously, interference tissues that only satisfy the spatial positional relationship but not the spatial structural relationship can be excluded. This makes it possible to accurately detect multiple feature points in a medical image.

[0104] Furthermore, according to the medical image processing apparatus and medical image processing method of the first embodiment, since the position score and structural score are independent of the detection process, only high detection accuracy is required for the detection model 1, and the position score and structural score are required to accurately reflect the spatial relationship between a target point and a reference point or between multiple target points. For this reason, compared to the conventional combined detection model which needs to consider the above factors simultaneously, the medical image processing apparatus and medical image processing method of this embodiment can obtain high detection accuracy for each target point while simultaneously ensuring accurate spatial positional relationships and spatial structural relationships. From these two angles, it is possible to ensure that detection results with high detection accuracy and accurate spatial relationships are obtained. As a result, multiple feature points in a medical image can be accurately detected.

[0105] (Second embodiment) Next, a medical image processing apparatus and a medical image processing method according to the second embodiment will be described based on Figures 8 to 10. In the following description, only the differences between the second embodiment and the first embodiment will be described, and the same configurations will not be explained.

[0106] The first embodiment is divided into a feature distribution calculation stage and a detection stage. In the feature distribution calculation stage, location features and structural features that follow a specific distribution are selected, and probability distributions corresponding to these location features and structural features are constructed to form location feature distributions and structural feature distributions, thereby scoring the location scores and structural scores of multiple candidate points. This method is also called a generative approach.

[0107] In contrast, the second embodiment employs a so-called discriminative approach as the basis for the positional and structural scores of multiple candidate points. Specifically, the second embodiment is divided into a classifier learning stage and a detection stage.

[0108] Figure 8 is a block diagram showing the functional configuration of a medical image processing apparatus according to the second embodiment. Figure 9 is a flowchart showing steps T300 to T400 of the classifier learning stage according to the second embodiment.

[0109] As shown in Figure 8, the medical image processing apparatus according to the second embodiment comprises a detection model 1, a reference point detection unit 10, a candidate point generation unit 20, a position score unit 30, a structure score unit 40, a classifier learning unit 70, and an output unit 60. The medical image processing apparatus according to the second embodiment may further include an update unit (details of which will be described later) which is not shown.

[0110] As shown in Figure 9, in step T300, the classifier is trained based on the location features.

[0111] Specifically, the classifier learning unit 70 sets target points, anomalies, and reference points in the medical images used as the training dataset 2. The classifier learning unit 70 then sets the target points in the training dataset 2 as positive samples and the anomalies as negative samples. Here, anomalies are points in the training dataset 2 whose visual features are similar to the target points but cannot be considered as target points. Next, the classifier learning unit 70 calculates the spatial positional relationship between the target points and the reference points, and the spatial positional relationship between the anomalies and the reference points, respectively, and, similar to T100 described above, selects the spatial positional relationship between the target points and the reference points as a positional feature. The positional feature classifier 3 is then trained by having the classifier learn this positional feature and the spatial positional relationship between the anomalies and the reference points.

[0112] Next, in step T400, the classifier is trained based on structural features.

[0113] Specifically, the classifier learning unit 70 sets target points and anomalies in the medical images, which serve as the training dataset 2. The classifier learning unit 70 then sets the target points in the training dataset 2 as positive samples and the anomalies as negative samples. Subsequently, the classifier learning unit 70 calculates the spatial structural relationships between multiple target points, the spatial structural relationships between multiple anomalies, and the spatial structural relationships between multiple target points and multiple anomalies in the training dataset 2. Similar to T200 described above, it selects the spatial structural relationships between multiple target points as structural features and trains the structural feature classifier 4 by having the classifier learn these structural features, the spatial structural relationships between multiple anomalies, and the spatial structural relationships between multiple target points and multiple anomalies.

[0114] Here, the positional feature classifier 3 and the structural feature classifier 4 can predict the classification probability of candidate points in the medical image input as input data, based on the target points and anomalies. Here, the classification probability of a candidate point is the probability that the candidate point is either a target point or an anomaly point.

[0115] Furthermore, for example, the classifier learning unit 70 may learn multiple weak classifiers with low correlation and perform combined prediction. Alternatively, the classifier learning unit 70 may simultaneously select as many effective features as possible to learn one strong classifier and perform prediction. The form and learning method of the classifier are not limited to those described above, and it is sufficient that it is learned based on the spatial positional relationship and spatial structural relationship of the target point or anomaly point, and that it can score candidate points for the target point.

[0116] Next, the detection step according to the second embodiment will be described with reference to Figure 10. Figure 10 is a flowchart showing steps S100 to S700 of the detection step according to the second embodiment. In the second embodiment, only the steps that differ from the first embodiment will be described, and steps that are the same as in the first embodiment will be denoted by the same reference numerals in Figure 10 and their description will be omitted. Examples of the same steps are steps S100, S200, S400, S600, and S700.

[0117] In step S200, the candidate point generation unit 20 generates multiple candidate points for the target point. The process then proceeds to step S310.

[0118] In step S310, the location scoring unit 30 assigns location scores to multiple candidate points based on the classification probabilities of the candidate points predicted by the location feature classifier 3. The process then proceeds to step S400.

[0119] Here, the position score of a candidate point represents the classification probability predicted by the position feature classifier 3 for that candidate point, i.e., the probability that the candidate point is a target point or a reference point. For example, the position score is greater than 0 and less than or equal to 1. The closer the position score is to 1, the higher the probability that the candidate point is a target point, and the closer the position score is to 0, the higher the probability that the candidate point is an anomaly point. Note that the position score may take other forms, as long as it can represent the classification probability of a target point or an anomaly point.

[0120] In step S400, the position score unit 30 selects candidate points whose position scores satisfy pre-set conditions and outputs them to the structure score unit 40. The process then proceeds to step S510.

[0121] In step S510, the structural scoring unit 40 assigns a structural score to multiple candidate point combinations based on the structural feature classifier 4. The process then proceeds to step S600.

[0122] Steps from step S600 onward are the same as in the first embodiment, so their explanation is omitted. This completes the flow of the second embodiment.

[0123] In the second embodiment, as in the first embodiment, the order of scoring the position score and the structure score may be reversed.

[0124] According to the medical image processing apparatus and medical image processing method of the second embodiment described above, a positional feature classifier is learned based on the spatial positional relationship (i.e., positional feature) between a target point and a reference point and the spatial positional relationship between an anomaly point and a reference point. A structural feature classifier is learned based on the spatial positional relationship (i.e., structural feature) between multiple target points, the spatial structural relationship between multiple anomaly points, and the spatial structural relationship between multiple target points and multiple anomaly points. Furthermore, based on the positional feature classifier and the structural feature classifier, positional scores and structural scores are scored for multiple candidate points of a single target point, and the classification probability that the candidate point belongs to the target point or the reference point is determined. Therefore, candidate points whose positional scores and structural scores satisfy pre-set conditions are selected as target points, and those target points are output as feature points. This makes it possible to accurately detect multiple feature points in a medical image.

[0125] Furthermore, according to the medical image processing apparatus and medical image processing method of the second embodiment, even when it is not possible to select a positional feature or structural feature that follows a specific distribution in a medical image, multiple feature points in the medical image can be accurately detected by distinguishing between target points and abnormal points at candidate points using a classifier.

[0126] (Third embodiment) Next, a medical image processing apparatus and a medical image processing method according to the third embodiment will be described based on Figure 11. Figure 11 is a flowchart of steps S100 to S700 of the detection stage according to the third embodiment. In the third embodiment, only steps that differ from the first embodiment will be described, and steps that are the same as in the first embodiment will be denoted by the same reference numerals in Figure 11 and their description will be omitted. Examples of the same steps are steps S100 and S700.

[0127] In step S220, the candidate point generation unit 20 generates multiple candidate points for the target point and outputs the visual prediction probability of the candidate points.

[0128] Using target point A as an example, the candidate point generation unit 20 generates multiple candidate points D of target point A. An Generate and each candidate point D An For the corresponding visual feature, the visual prediction probability Prob An The candidate point D An The output is the visual feature similarity. Here, the visual features include information extracted from the medical image, such as brightness, edge morphology, and patterns. Next, the process proceeds to step S320.

[0129] In step S320, the position scoring unit 30 assigns individual overall scores to multiple candidate points based on their positional features and visual features.

[0130] Specifically, the position score unit 30 calculates positional features and positional feature distributions, similar to the first embodiment, and performs positional evaluation on candidate points. For example, based on positional features F1, F2, and F3, the position score unit 30 determines candidate point D An Location score S DAn _ F1 S DAn _ F2 S DAn _ F3 Each of these is calculated. Next, the position score unit 30 calculates the position score S of each position feature according to the following equation (4). DAn _F1 S DAn _ F2 S DAn _ F3 and visual feature similarity Prob An By assigning different weighting coefficients to each point and adding the position scores weighted by each weighting coefficient, the candidate point D is obtained. An Individual overall score S An Calculate.

[0131]

number

[0132] Here, λ represents the weighting coefficient. For example, the position score unit 30 assigns a higher weighting coefficient to positional features or visual features with a higher confidence level. Next, the process proceeds to step S420.

[0133] In step S420, the position score unit 30 selects candidate points whose individual overall scores satisfy pre-set conditions and outputs them to the structure score unit 40. The specific selection method is the same as in the first embodiment, so the explanation is omitted. Next, the process proceeds to step S520.

[0134] In step S520, the structural scoring unit 40 scores a combined overall score for each candidate point combination based on the structural characteristics and the individual overall scores of each candidate point.

[0135] Taking target points A, B, and C as an example, in step S220, the candidate point generation unit 20 generates multiple candidate points D of target points A, B, and C. An , D Bn , D Cn If a result is generated, the structural score unit 40 will determine the candidate point combination n(D) based on the structural features F4, F5, and F6. An ,D Bn ,D Cn ) Structural score S n _ F1 S n _ F2 S n _ F3The position score unit 30 calculates the candidate point D. An , D Bn , D Cn Each individual overall score S An S Bn S Cn Assuming that the following has been calculated, the structural score section 40 is calculated by the following equation (5) as follows: candidate point combination n(D An ,D Bn ,D Cn Calculate the overall score Sn for the combination of ).

[0136]

number

[0137] Here, λ represents the weighting coefficient. Next, we proceed to step S620.

[0138] In step S620, the structural score unit 40 selects candidate points included in candidate point combinations whose combined overall score satisfies pre-set conditions and outputs them to the output unit 60. The specific selection method is the same as in the first embodiment, so the explanation is omitted. Next, the process proceeds to step S700.

[0139] In step S700, the output unit 60 outputs the target point as a feature point. This completes the flow of the third embodiment.

[0140] In addition, in the third embodiment, as in the first embodiment, the order of scoring the position score and the structure score may be reversed.

[0141] According to the medical image processing apparatus and medical image processing method of the third embodiment described above, based on the visual features in the medical image, an individual overall score can be assigned to candidate points and a combined overall score can be assigned to combinations of candidate points by combining visual feature similarity, which indicates the visual prediction probability, with positional features and structural features. This makes it possible to accurately detect multiple feature points in a medical image.

[0142] For example, in medical images of postoperative subjects, the positional feature score of interference points may be higher than the positional feature score of target points. Even in such cases, by using visual feature similarity to assign an overall score to interference points with high positional scores, interference points can be eliminated, and multiple feature points in medical images can be accurately detected.

[0143] (modified version) The embodiments described above are merely examples, and the medical image processing apparatus and medical image processing method according to this application are not limited to the configurations described in each embodiment.

[0144] For example, in each embodiment, the medical image processing device may further include an update unit that updates the detection model used to calculate the location feature distribution and the structural feature distribution, and the detection model used in the learned location feature classifier and structural feature classifier, through online learning. This makes it possible to obtain location scores and structural scores that are more adapted to the current dataset.

[0145] Furthermore, the target points, candidate points, reference points, etc., used in the above description are merely examples given to illustrate the image processing procedure according to this embodiment, and are not intended to limit their position, format, number, etc. Their specific positions, formats, number, etc., may be any as long as they enable the realization of the medical image processing apparatus and medical image processing method according to this embodiment.

[0146] Furthermore, the reference point detection unit 10, candidate point generation unit 20, position score unit 30, structure score unit 40, feature distribution calculation unit 50, output unit 60, and classifier learning unit 70 in each embodiment are realized by a processing circuit such as a processor. In this case, the processing functions of each of these processing units are stored in a memory unit in the form of a program that can be executed by a computer. The processing circuit then reads and executes each program stored in the memory unit to realize the processing function corresponding to each program. In other words, the medical image processing device in each embodiment has the functional configuration shown in Figures 1 and 8 when the processing circuit has read each program. In this case, the processing of each step shown in Figures 2, 3, 5 to 7, and 9 to 11 is realized, for example, when the processing circuit reads and executes the program corresponding to each processing function from the memory unit.

[0147] Furthermore, in each embodiment, the processing circuit is not limited to being implemented by a single processor, but may be composed of a combination of multiple independent processors, with each processor executing a program to realize each processing function. Also, the processing functions of each processing unit may be implemented by appropriately distributing or integrating them across one or more processing circuits. Moreover, the processing functions of each processing unit may be implemented by hardware such as circuits only, software only, or a mixture of hardware and software. In addition, although an example in which the program corresponding to each processing function is stored in a single memory circuit has been described here, the embodiments are not limited to this. For example, the program corresponding to each processing function may be distributed and stored in multiple memory circuits, and each processing unit may read and execute each program from each memory circuit.

[0148] Furthermore, the term "processor" used in the above-described embodiments refers to circuits such as a CPU (Central Processing Unit), a GPU (Graphics Processing Unit), an Application Specific Integrated Circuit (ASIC), or a programmable logic device (e.g., a Simple Programmable Logic Device (SPLD), a Complex Programmable Logic Device (CPLD), and a Field Programmable Gate Array (FPGA)). Here, instead of storing the program in a memory circuit, the processor may be configured to directly incorporate the program into its circuitry. In this case, the processor realizes its function by reading and executing the program incorporated into the circuitry. Moreover, each processor in this embodiment is not limited to being configured as a single circuit; multiple independent circuits may be combined to form a single processor, and its function may be realized in this way.

[0149] Here, the program executed by the processor is provided pre-installed in ROM (Read Only Memory) or memory circuits. Alternatively, this program may be provided as a file in an installable or executable format on a computer-readable, non-transient storage medium such as a CD (Compact Disk)-ROM, FD (Flexible Disk), CD-R (Recordable), or DVD (Digital Versatile Disk). Furthermore, this program may be stored on a computer connected to a network such as the Internet and provided or distributed by downloading it via the network. For example, this program consists of modules containing the processing functions described above. In actual hardware, the CPU reads the program from a storage medium such as ROM and executes it, loading each module onto the main memory and generating it in the main memory.

[0150] Furthermore, in the embodiments described above, each component of each illustrated device is a functional concept and does not necessarily have to be physically configured as shown. In other words, the specific form of distribution or integration of each device is not limited to that shown, and all or part of them can be functionally or physically distributed or integrated in any unit according to various loads and usage conditions. Moreover, each processing function performed by each device can be implemented, in whole or in any part, by a CPU and a program that is analyzed and executed by the CPU, or by hardware using wired logic.

[0151] Furthermore, among the processes described in the embodiments described above, all or part of the processes described as being performed automatically can be performed manually, or all or part of the processes described as being performed manually can be performed automatically by known methods. In addition, the processing procedures, control procedures, specific names, and information including various data and parameters shown in the above document and drawings can be arbitrarily changed unless otherwise specified.

[0152] The various types of data discussed in this specification are typically digital data.

[0153] According to at least one embodiment described above, multiple feature points in a medical image can be accurately detected.

[0154] While several embodiments have been described, these embodiments are presented as examples only and are not intended to limit the scope of the invention. These embodiments can be implemented in a variety of other forms, and various omissions, substitutions, modifications, and combinations of embodiments are possible without departing from the spirit of the invention. These embodiments and their variations are included in the scope and spirit of the invention, as well as in the claims and their equivalents. [Explanation of Symbols]

[0155] 1. Detection Model 2. Training dataset 3. Location Feature Classifier 4. Structural Feature Classifier 10 Reference point detection unit 20 Candidate point generation section 30 Position Score Section 40 Structural score section 50 Feature Distribution Calculation Unit 60 Output section 70 Classifier Learning Unit

Claims

1. A medical image processing apparatus for detecting a plurality of feature points in a medical image, comprising: a reference point detection unit that detects, for each of a plurality of target points corresponding to the plurality of feature points, a reference point that has spatial correlation with the target point in the medical image; a candidate point generation unit that generates, for each of the plurality of target points, a plurality of candidate points corresponding to the target points by using a detection model; a position score unit that assigns a position score to the candidate point based on a position feature indicating a spatial positional relationship between the target point and the reference point in the medical image for each of the plurality of target points, and selects a candidate point whose position score satisfies a preset condition; a structure score unit that assigns a structure score to each of a plurality of candidate point combinations obtained by combining the candidate points among the plurality of target points based on a structural feature that indicates a spatial structural relationship between the plurality of target points in the medical image, thereby improving a detection rate of the plurality of feature points in the medical image and selecting a candidate point combination whose structure score satisfies a preset condition; an output unit that outputs the plurality of feature points in the medical image via a display unit based on the candidate points selected by the position score unit and the candidate point combinations selected by the structure score unit; A medical image processing device comprising:

2. the position score unit selects a plurality of candidate points whose position scores satisfy the condition from the plurality of candidate points generated by the candidate point generation unit; the structure score unit selects a candidate point combination whose structure score satisfies the condition from a plurality of candidate point combinations obtained by combining the plurality of candidate points selected by the position score unit; the output unit outputs a plurality of candidate points included in the candidate point combination selected by the structure score unit as the plurality of feature points. The medical image processing device according to claim 1 .

3. the structure score unit selects a candidate point combination whose structure score satisfies the condition from a plurality of candidate point combinations obtained by combining a plurality of candidate points generated by the candidate point generation unit, the position score unit selects a plurality of candidate points whose position scores satisfy the condition from a plurality of candidate points included in the candidate point combination selected by the structure score unit; The output unit outputs the plurality of candidate points selected by the position score unit as the plurality of feature points. The medical image processing device according to claim 1 .

4. a feature distribution calculation unit that calculates a position feature distribution indicating a probability distribution of the position feature and a structure feature distribution indicating a probability distribution of the structure feature based on a learning data set of the detection model, The position score unit assigns the position score to the candidate point based on the position feature and the position feature distribution corresponding to the position feature; the structure score unit assigns the structure score to the candidate point combination based on the structure feature and the structure feature distribution corresponding to the structure feature; 4. The medical image processing apparatus according to claim 1.

5. The method further includes a feature distribution calculation unit that calculates a binding position feature distribution indicating a probability distribution of a binding position feature obtained by combining a plurality of spatial positional relationships as the position feature, and a binding structure feature distribution indicating a probability distribution of a binding structure feature obtained by combining a plurality of structure features as the structure feature, based on a learning data set of the detection model, the position score unit assigns the position score to the candidate point based on the binding position feature and the binding position feature distribution corresponding to the binding position feature; the structure score unit assigns the structure score to the candidate point combination based on the bond structure feature and the bond structure feature distribution corresponding to the bond structure feature; 4. The medical image processing apparatus according to claim 1.

6. A classifier training unit is further provided which trains a position feature classifier based on the position features and a spatial positional relationship between an abnormal point in the medical image and the reference point, and trains a structure feature classifier based on the structure features, a spatial structural relationship between a plurality of abnormal points in the medical image, and a spatial structural relationship between the plurality of target points and the plurality of abnormal points in the medical image; The position score unit assigns the position score to the candidate point based on the position feature classifier; The structure score unit assigns the structure score to the candidate point combination based on the structure feature classifier.

4. The medical image processing apparatus according to claim 1.

7. The candidate point generation unit further generates visual features of the plurality of candidate points for each of the plurality of target points using the detection model; The position score unit assigns an individual overall score as the position score to the candidate point based on the position feature and the visual feature; the structure score unit assigns a combination total score as the structure score to the candidate point combination based on the structural feature and the individual total score; 4. The medical image processing apparatus according to claim 1.

8. an update unit configured to update the position feature distribution and the structure feature distribution by online learning based on the position score and the structure score, The medical image processing device according to claim 4 .

9. an update unit configured to update the position feature classifier and the structure feature classifier by online learning based on the position score and the structure score, The medical image processing device according to claim 6 .

10. 1. A medical image processing method for detecting a plurality of feature points in a medical image, comprising: a reference point detection step of detecting, for each of a plurality of target points corresponding to the plurality of feature points, a reference point having a spatial correlation with the target point in the medical image; a candidate point generating step of generating, for each of the plurality of target points, a plurality of candidate points corresponding to the target points by using a detection model; a position score step of assigning a position score to the candidate point based on a position feature indicating a spatial positional relationship between the target point and the reference point in the medical image for each of the plurality of target points, and selecting a candidate point whose position score satisfies a preset condition; a structure score step of selecting a candidate point combination whose structure score satisfies a preset condition by assigning a structure score to each of a plurality of candidate point combinations obtained by combining the candidate points among the plurality of target points based on a structural feature indicating a spatial structural relationship between the plurality of target points in the medical image, thereby improving a detection rate of the plurality of feature points in the medical image; and an output step of outputting the plurality of feature points in the medical image via a display unit based on the candidate points selected in the position score step and the candidate point combinations selected in the structure score step; A medical image processing method comprising:

11. the position score step selects a plurality of candidate points whose position scores satisfy the condition from the plurality of candidate points generated in the candidate point generation step; the structure score step selects, from the plurality of candidate points selected in the position score step, a combination of candidate points whose structure score satisfies the condition; the output step outputs a plurality of candidate points included in the candidate point combination selected in the structure score step as the plurality of feature points. The medical image processing method according to claim 10.

12. the structure score step selects a candidate point combination whose structure score satisfies the condition from the plurality of candidate points generated in the candidate point generation step; the position score step selects a plurality of candidate points whose position scores satisfy the condition from a plurality of candidate points included in the candidate point combination selected in the structure score step; The output step outputs the plurality of candidate points selected in the position score step as the plurality of feature points. The medical image processing method according to claim 10.

13. The method further includes a feature distribution calculation step of calculating a position feature distribution indicating a probability distribution of the position feature and a structure feature distribution indicating a probability distribution of the structure feature based on a learning data set of the detection model, The position score step assigns the position score to the candidate point based on the position feature and the position feature distribution corresponding to the position feature; the structure score step assigns the structure score to the candidate point combination based on the structure feature and the structure feature distribution corresponding to the structure feature; The medical image processing method according to any one of claims 10 to 12.

Images