KR20220164097A - Apparatus, method and computer program for anayzing medical image using classification and segmentation - Google Patents

Abstract

The present invention relates to a method for analyzing a medical image by an analysis apparatus, which comprises: a step a of receiving a target image to detect an abnormal part; a step b of extracting a region of interest from the target image; a step c of using a machine learning model learned from normal determination data and abnormal determination data corresponding to the region of interest as learning data to classify an abnormal state of the region of interest; a step d of segmenting the abnormal region in the region of interest; and a step e of using a classification value corresponding to the abnormal state calculated in the step c and an abnormal probability value for each pixel of the abnormal region calculated in the step d to determine the abnormal state of the target image or region of interest. Accordingly, the method can quickly and conveniently read whether an abnormal region exists in a medical image.

Description

Medical image analysis method, apparatus and computer program using classification and segmentation

The present invention relates to a method, apparatus, and computer program for analyzing a medical image, and more particularly, to a method, apparatus, and computer program for analyzing abnormal parts and states included in a medical image using classification and segmentation.

Currently, the problem of lack of doctors in Korea is getting worse, and in particular, there is a problem in which rapid response is not made due to the lack of specialists in urgent medical fields. Musculoskeletal abnormalities such as fractures account for the largest number of patients who come to the emergency room, but in many cases, a musculoskeletal specialist or a radiologist is absent in the emergency room. Plain radiographs play a very important role in diagnosing fractures in patients with initial trauma, but up to 80% of initial diagnosis failure rates due to simple radiograph reading errors in emergency rooms have been reported, especially with phalanges imaging. It is reported that the detection of abnormalities such as fractures for . If the reading by the radiologist or musculoskeletal specialist is not accurately performed in real time, the patient's stay in the emergency room becomes longer, and problems such as serious fracture sequelae due to misdiagnosis, patient complaints, and lowered reliability of medical institutions may occur. So, we need to solve this. In particular, some fractures have a low prevalence, but are very difficult to detect, and if not detected early, they can cause problems such as osteonecrosis and cause serious sequelae.

Therefore, attempts have been made to automatically read radiographs using machine learning models in the same technology field, but the method using machine learning models has limitations in accuracy, and due to the nature of machine learning called black box, automatic Due to insufficient explanation of the analysis results, it is not used well in the field.

An object of the present invention is to solve the above problems, and to quickly and conveniently read whether or not there is an abnormal part in a medical image.

Another object of the present invention is to improve the classification accuracy of an abnormal state and provide precise analysis for each region in medical image analysis using a deep learning convolutional neural network technique.

In addition, the present invention provides a medical image analysis method, apparatus, and program that overcome the limitations of conventional image analysis by presenting characteristics of analysis results and providing evidence through visualization in medical image analysis using machine learning. to do for a different purpose.

In order to achieve this object, the present invention provides a method for analyzing a medical image by an analysis device, comprising step a of receiving a target image to detect an abnormal part, step b of extracting a region of interest from the target image, and corresponding to the region of interest. Step c of classifying the abnormal state of the region of interest using the machine learning model trained with the normal discrimination data and the abnormal discrimination data as learning data, Step d of segmenting the abnormal part in the region of interest and step e of determining the abnormal state of the target image or region of interest using the classification value corresponding to the detected abnormal state and the abnormal probability value for each pixel of the abnormal region calculated in step d.

In addition, the present invention is an apparatus for analyzing a medical image, an input unit for receiving a target image to detect an anomaly, a region of interest extractor for extracting a region of interest from the target image, normal determination data corresponding to the region of interest, and an abnormality An anomaly classification unit that classifies whether or not an abnormal area exists in the region of interest using a machine learning model learned from discrimination data as learning data, an anomaly region detection unit that segments the abnormal region in the region of interest, and an anomaly classification unit It is characterized by including an integrated analysis unit that determines the abnormal state of the region of interest by using the calculated classification value corresponding to the presence or absence of the abnormal region and the abnormal probability value for each pixel of the abnormal region calculated by the abnormal region detection unit.

According to the present invention as described above, it is possible to quickly and conveniently read whether an abnormal part exists in a medical image.

In addition, according to the present invention, in medical image analysis using a deep learning convolutional neural network technique, it is possible to improve the classification accuracy of an abnormal state and provide precise analysis for each region.

In addition, in medical image analysis using machine learning, the present invention can overcome the limitations of conventional image analysis by presenting characteristics of analysis results and providing evidence through visualization.

1 is a block diagram for explaining the configuration of a medical image analysis apparatus according to an embodiment of the present invention;
2 is a diagram for explaining a classification value input to an integrated analysis unit according to an embodiment of the present invention;
3 is a view for explaining detection and visualization of abnormal parts according to an embodiment of the present invention;
4 is a flowchart for explaining a medical image analysis method according to an embodiment of the present invention;
5 is a flowchart for specifically explaining a method for determining an abnormal state according to an embodiment of the present invention;
6 is a flowchart for explaining in detail a method for determining an abnormal state according to another embodiment of the present invention.

The above objects, features and advantages will be described later in detail with reference to the accompanying drawings, and accordingly, those skilled in the art to which the present invention belongs will be able to easily implement the technical spirit of the present invention. In describing the present invention, if it is determined that the detailed description of the known technology related to the present invention may unnecessarily obscure the subject matter of the present invention, the detailed description will be omitted. Hereinafter, preferred embodiments according to the present invention will be described in detail with reference to the accompanying drawings. In the drawings, the same reference numerals are used to indicate the same or similar elements, and all combinations described in the specification and claims may be combined in any manner. And unless otherwise specified, it should be understood that references to the singular may include one or more, and references to the singular may also include plural.

The present invention is applicable to various fields. In this specification, the contents of the technology will be described focusing on the determination of abnormal conditions in the musculoskeletal system, but the present invention is not limited to the types of abnormal diseases, and is applicable to various medical image analysis to determine abnormal conditions such as cardiovascular diseases and cerebrovascular diseases. can be utilized

On the other hand, in this specification, the term 'abnormal part' refers not only to a part judged to be abnormal in the binary classification of normal and abnormal, but also to various abnormal states (suspicion, suspicion, tumor, fracture, fracture nonunion, acute fracture, etc.)

In addition, the term 'abnormal state' used herein may be understood as a term meaning a state of an abnormal part existing in a target image or region of interest. The 'abnormal state' may include normal or abnormal, and more specifically, may include specific conditions of abnormal parts such as normal, suspected, cyst, benign tumor, negative tumor, and malignant tumor. Alternatively, when an image to be analyzed is an image suspected of having a fracture, the 'abnormal state' may include states such as normal, suspected fracture, fracture non-union, and acute fracture.

Therefore, the classification value corresponding to the abnormal state may also be defined differently according to the classification method. For example, when the anomaly classification unit 150 performs binary classification on a target image or region of interest, an image classified as a normal state according to the binary classification includes a classification value corresponding to a non-'abnormal state', an anomaly An image classified as a state will have a classification value corresponding to 'abnormal state'. If the abnormal classification unit 150 multi-classifies a target image or region of interest, an image classified as a normal state has a classification value corresponding to a non-'abnormal state', and an image classified as a suspicious state has an 'abnormal state' The type of is a classification value corresponding to the 'suspicious' category, the image classified as having a cyst is a classification value corresponding to the 'abnormal state' category, and the image classified as having a tumor is 'abnormal' The type of 'status' may have a classification value corresponding to the 'tumor' category. That is, an abnormal state is a concept that includes a plurality of state types, and can be understood as a concept representing various states of an abnormal region, a region of interest, and a target image, rather than simply a concept opposite to a normal state.

Hereinafter, a medical image analysis apparatus according to an embodiment of the present invention will be described with reference to FIGS. 1 to 3 .

Referring to FIG. 1 , the medical image analysis apparatus 100 according to an embodiment of the present invention includes an input unit 110, an ROI extraction unit 130, an anomaly classification unit 150, an anomaly part detection unit 160, and an integration unit. It includes an analysis unit 170, and may further include a learning unit 120 and an output unit 190.

The input unit 110 may be an input module that receives a target image from a user, or may be a communication module that receives a target image from another electronic device (not shown) through a wired or wireless communication network. In another embodiment, the input unit 110 may be an imaging module that acquires an image of an object. The target image may be a 2D or 3D color or black and white image, and may be obtained through an imaging device such as a camera, a radiation imaging device, a computed tomography device, or a capture function of an electronic device. The input unit 110 may transmit the received target image to the ROI extractor 130 . If the medical image analysis device 100 according to an embodiment of the present invention is a device that analyzes fracture images professionally, the input unit 110 may receive a fracture image.

The ROI extractor 130 extracts one or more ROIs from the target image. Here, the region of interest refers to a specific region having importance in analyzing a target image. In machine learning analysis, the importance of a specific region is often based on domain knowledge to which the image belongs. That is, the analysis based on the region of interest is intended to increase accuracy by removing noise in regions other than the main region of interest that may cause errors in analysis. For example, when a medical image as shown in FIG. 3 is received as a target image (a radiation image excluding a heat map is received), and it is desired to analyze whether or not a wrist region is fractured, the medical image analysis apparatus 100 uses the wrist When determining the abnormal state of a part, the abnormal state can be determined by extracting parts such as the radius, ulna, and navicular bone that require confirmation as a region of interest.

The region of interest extractor 130 may manually or automatically detect the region of interest. For example, a region corresponding to location information set by a user may be extracted as the region of interest. That is, in the wrist image of FIG. 2 , the user may directly set one region such as the radius, ulna, or navicular bone, and the ROI extractor 130 extracts the region designated by the user as the ROI, ) and the abnormal part detection unit 160.

As another embodiment, the ROI extractor 130 may extract a region corresponding to the specific region as the ROI by applying an automatic detection model learned for the specific region to the query image. Here, the specific region means a region that is mainly considered for bone age measurement in the above-described example. The automatic detection unit 137 may include an automatic detection model, which is a deep neural network learned (trained and evaluated) using a plurality of images of a specific region in order to extract a specific image from all images. As the automatic detection model, deep neural network-based Convolutional Neural Network (CNN), Faster-RCNN, and Yolo detection techniques may be used, but are not limited thereto.

For example, the learning unit 120 may train an automatic detection model using a plurality of radial images, and in this case, the ROI extractor 130 may use an arbitrary target image (including the wrist) in the learned automatic detection model. A region corresponding to the radius may be extracted as a region of interest by applying a hand bone image).

The learning unit 120 may independently learn one or more automatic detection models and provide them to the ROI extractor 130 . Each automatic detection model is a machine learning model, and may be a deep neural network or regression analysis model composed of one or more layers.

The learning unit 120 may train an automatic detection model using a plurality of ROI images previously identified for each ROI, that is, classified according to a preset identification value, as training data. For example, the learning unit 120 may train the automatic detection model by using regions of interest classified according to the frequency of fractures for each region of interest, difficulty of detection, degree of sequelae when not detected, and the like as learning data. For example, the learning unit 120 may count the frequency of fracture occurrence for each part in a plurality of medical image data in which fracture occurs, and may train an automatic detection model for each part where the fracture frequency is high. Alternatively, the learning unit 120 may train the automatic detection model by receiving medical information about the difficulty of detection or the degree of aftereffects for each region of interest from the user as a learning data identification value.

The learning unit 120 may learn not only the automatic detection model, but also the machine learning model used in the anomaly classification unit 150 based on the normal discrimination data and the anomaly discrimination data corresponding to the region of interest, and the anomaly part detection unit 160 ) can also train the segmented CNN model used in

In addition, the learning unit 120 is integrated analysis model used in the integrated analysis unit 170

The ROI extractor 130 extracts one or more ROI (ROI 1, 2, ..., N) from the target image and delivers the extracted ROI to the anomaly classifier 150 and the anomaly detection unit 160. can The extraction of this region of interest is intended to improve accuracy. Instead of analyzing the image using the entire image as input data, as in the conventional general image analysis method, machine learning modeling and application are performed for each major detailed region. to enable precise analysis.

The anomaly classification unit 150 determines the abnormal state of the region of interest according to whether there is an abnormal part in the region of interest by using a machine learning model learned from normal determination data and abnormality determination data corresponding to the region of interest as learning data. can be classified. Existence of an abnormal region in the region of interest means that there is an abnormal region in the target image as a result, so it can be considered that the abnormality classification unit 150 determines and classifies the normal/abnormal state of the target image. For example, when the target image is a lung image and the machine learning model of the abnormal classification unit 150 is a machine learning model learned from abnormal discrimination data in which a pulmonary nodule or tumor is present and normal discrimination data in which no pulmonary nodule or tumor is present. , The anomaly classification unit 150 may determine whether there is a pulmonary nodule or a tumor in the received lung image.

As another embodiment, when the target image is a brain image and the machine learning model of the anomaly classification unit 150 is a machine learning model learned from abnormal discrimination data with cerebral hemorrhage and normal discrimination data without brain hemorrhage, the anomaly classification unit 150 may determine whether there is brain hemorrhage in the received brain image, and accordingly, an abnormal state of the target image or an abnormal state of the region of interest may be classified.

The machine learning model used in the anomaly classification unit 150 is a deep neural network composed of one or more layers, a plurality of convolution layers that create a feature map for features of the region of interest, and a plurality of It may be a convolutional neural network (CNN) including a pooling layer performing subsampling between convolutional layers. A convolutional neural network can extract features from an input image by alternately performing convolution and subsampling on the input image. A feature or feature value refers to a vector value representing key features important in classifying or identifying an image. According to an embodiment of the present invention, the first feature input to the integrated analysis unit 170 may be a value output from one layer or a value input to one layer among one or more layers constituting the deep neural network.

The convolutional neural network includes several convolution layers, several subsampling layers (subsampling layer, Max-Pooling layer, and Pooling layer), a Global Average Pooling layer (GAP layer), and a fully-connected layer (Fully-Pooling layer). Connected layer), Softmax layer, and the like. The convolution layer is a layer that performs convolution on the input image, and the subsampling layer is a layer that extracts the maximum or average value locally from the input image and maps it to a 2D image. It enlarges the local area and performs subsampling. can be done

The convolution layer requires information such as the size of the kernel, the number of kernels to be used (the number of maps to be created), and a table of weights to be applied during the convolution operation. In the subsampling layer, information on the size of a kernel to be subsampled and whether to select a maximum value or a minimum value among values within a kernel area is required.

In FIG. 2 showing an example of a convolutional neural network (CNN), the convolutional neural network includes an integration layer (layer 2) in which subsampling is performed between convolutional layers (layer 1 and layer 3), and a GAP layer at the end, It may include a fully connected layer and a softmax layer. This is for convenience of description, and a convolutional neural network can be constructed by combining layers having different characteristics from those applied in this example in various ways, and the present invention is not limited by the configuration and structure of a deep neural network.

The classification value output from the anomaly classification unit 150 is a value calculated by applying the region of interest to the deep neural network machine learning model, and as shown in FIG. 2, the input value, output value, or input value of one layer included in the deep neural network. It may include at least one of a value or a calculation value calculated using an output value. As described above, the classification value may be a value corresponding to normal/abnormal according to binary classification, and normal/suspicious/tumor/tumor (negative, benign, malignant) or normal/fracture according to multiple classifications. It may be a value corresponding to a class according to user settings such as suspected/fracture nonunion/acute fracture/fracture.

The abnormal part detection unit 160 may segment the abnormal part in the region of interest, and an abnormality probability value for each pixel may be calculated as a result of the abnormal part detection unit 160 . It can be understood that the abnormal probability value means a value corresponding to a probability that is determined to be abnormal (abnormal).

The anomaly detection unit 160 may predict the labels of all pixels and classify all pixels in the target image, that is, the region of interest into a pre-specified number of classes (dense prediction). For example, the abnormal part detection unit 160 may classify and predict the binary classification of normal and abnormal in units of pixels, and present a probability of determining abnormality. Based on this, the visualization unit 180 may visualize the characteristics of the abnormal part. In detail, the visualization unit 180 may display a color or image corresponding to the abnormality probability value on the target image by using the abnormality probability value for each pixel received from the abnormality detection unit 160 .

For example, in the wrist image shown in FIG. 3 , regions with a high abnormality probability value may be displayed in red, and regions with a low abnormality probability value may be displayed in blue. In this case, a heat map as shown in FIG. 3 this can be created. The user can intuitively identify the abnormal part through the heat map, and check the probabilistic possibility of a disease such as a fracture for each major area (region of interest). On the other hand, the criterion of low or high odds probability value may vary according to a threshold set in advance by a user, which may be adjusted according to a user setting.

As another embodiment, when the target image is a brain image, the abnormal part detection unit 160 may attempt to distinguish between normal and abnormal (cerebral hemorrhage) for each region (pixel). The abnormal part detection unit 160 learns the abnormal region (pixels) in which brain hemorrhage is present and the normal region (pixels) in which brain hemorrhage is not present in each pixel unit, and converts the abnormal probability value of brain hemorrhage in the received brain image into pixel units. can present

The integrated analysis unit 170 uses the classification value corresponding to the abnormal state calculated by the abnormal classification unit 150 and the abnormality probability value for each pixel of the abnormal region calculated by the abnormal region detection unit 160 to determine an abnormality in the target image or region of interest. status can be judged. The classification value used in the integrated analysis unit 170 is an automatic classification feature for the abnormal state of the entire image of the target region or the region of interest, and the abnormal probability value is the abnormal probability value for each pixel calculated by determining the abnormal state in units of pixels (characteristic value) is. The present invention has characteristics differentiated from those of the prior art in that the abnormal state of the region of interest is determined by comprehensively considering the classification value and the abnormal probability value.

라 하면, 각 특징값 M^k 의 평균값인 A^k는 다음과 같은 수학식으로 표현할 수 있다.For example, the anomaly classification unit 150 according to an embodiment of the present invention converts the value of a feature map prior to delivery to the final softmax layer included in the deep neural network to the integrated analysis unit 170. ) can be used as an input. That is, the value of the feature map may be a classification value calculated by the anomaly classification unit 150 . In the present embodiment, in the example of FIG. 2 in which the average value of each feature map is obtained through Global Average Pooling (GAP) and transferred to the softmax layer through a Fully Connected Layer (FC Layer), the anomaly classification unit 150 As a classification value, the mean value vector of each feature map calculated through GAP can be used. Then, by passing this as an input to the integrated analysis unit 170, the classification value of the region of interest may be applied to the integrated analysis model like classification values of other regions of interest. In this case, the classification value of the region of interest n can be expressed as Fn = [A ¹ , A ² A ³ ,..., A ^k ] (k is the total number of channels), and the position of (i,j) in k channel feature value

, A ^k , which is an average value of each characteristic value M ^k , can be expressed by the following equation.

As another embodiment, the anomaly classification unit 150 may use scores or normalized probability values of each class for classifying the final class of the deep neural network as a classification value. For example, each class score value of the softmax layer or a normalized probability value obtained as a result of the softmax layer may be calculated as a classification value of the corresponding region of interest and input to the integrated analysis unit 170 .

(M= 클래스의 총 개수)이 되며, 통합분석부(170)로는 관심 영역별 분류값

이 입력될 수 있다. S ^C means the score of class C, and w ^c _i means the weight of A ⁱ . In this case, in this case, the classification value of region of interest n

(M = the total number of classes), and the integrated analysis unit 170 is a classification value for each region of interest.

can be entered.

In this way, the classification value for each region of interest calculated by the abnormal classification unit 150 is input to the integrated analysis unit 170, and the integrated analysis unit 170 uses the classification value to finally determine the abnormal state of the target image or the region of interest. can judge

As shown in FIG. 4 , the integrated analysis unit 170 determines the abnormal state of the region of interest by applying the classification value and the abnormal probability value calculated by the abnormal classification unit 150 and the abnormal part detection unit 160 to the integrated analysis model. Here, the integrated analysis model may be a neural network or a regression model learned using classification values corresponding to abnormal states of a plurality of training images and abnormal probability values for each pixel in the abnormal region as learning data.

As another embodiment, the integrated analysis unit 170 counts the number of first pixels whose abnormality probability value exceeds a preset first threshold, calculates a ratio of the number of first pixels in the target image or region of interest, and calculates If any one of the ratio and the classification value exceeds the preset second threshold, it may be determined that an abnormal part exists in the target image or region of interest.

As another embodiment, the integrated analysis unit 170 counts the number of pixels whose abnormality probability value exceeds a first threshold, and applies the abnormality probability value of the pixels exceeding the first threshold to a preset function to determine the target image or region of interest. An abnormal state can be judged.

Equation 3 above shows an example of a function for determining the abnormal state of the region of interest n in the integrated analysis unit 170. In the above equation, T denotes a predetermined threshold value, p ⁱ denotes an odd probability value of a pixel, and i denotes position information of a pixel. When the abnormal state is classified into a plurality of classes, the abnormal probability value p ⁱ of pixel i may be replaced with the probability p_c ⁱ that pixel i corresponds to a specific abnormal state (class). A function applied to an abnormal probability value of pixels exceeding a threshold value may vary according to a user setting. For example, the user may calculate an average value of abnormal probability values of pixels exceeding a threshold value as a feature value of a corresponding region of interest, and the calculated feature value may be a value corresponding to an abnormal state.

In another embodiment, the integrated analysis unit 170 may adjust the first threshold according to a classification value, that is, an abnormal state of the target image or region of interest. For example, when the anomaly classification unit 150 classifies the target image as having an abnormal region in the target image, the integrated analysis unit 170 may reduce the false negative by lowering or raising the first threshold value from the reference value. .

In another embodiment, the integrated analysis unit 170 may determine whether to use the abnormality probability value according to the classification value calculated by the abnormality classification unit 150 in a decision tree method. If it is determined that the abnormal probability value is used, the integrated analysis unit 170 may determine the abnormal state of the region of interest using the number of pixels having a preset probability value. When it is determined that the integrated analysis unit 170 also utilizes the abnormal probability value, as described above, the abnormal probability value itself of pixels having an abnormal probability value exceeding a predetermined threshold is calculated, or the number of pixels having an abnormal probability value exceeding the threshold value is exceeded. An abnormal state of the region of interest may be determined using

Finally, the output unit 190 may output at least one of the results calculated by the abnormal classification unit 150 , the abnormal part detection unit 160 , and the integrated analysis unit 170 . That is, the output unit 190 calculates the classification value calculated by the abnormal classification unit 150, the abnormal probability value or abnormal probability value for each pixel calculated by the abnormal part detection unit 160, or the representative value calculated by the integrated analysis unit 170. At least one of result values for whether or not the target image or the region of interest calculated in is included may be provided.

The medical image analysis apparatus 100 of the present invention is conventional image analysis in that it performs analysis by integrating classification values of normal or abnormal states for each region of interest and pixel-specific information of an abnormal region calculated through segmentation. High analysis accuracy that the analysis method did not provide can be provided.

4 is a flowchart illustrating a method for analyzing a medical image according to an embodiment of the present invention.

Referring to FIG. 4 , the electronic device according to an embodiment of the present invention may receive a target image (S100) and extract one or more regions of interest from the target image (S200).

In step 200, the electronic device can extract the region corresponding to the location information preset by the user as the region of interest, and apply the automatic detection model learned for the specific region to the query image to determine the region corresponding to the specific region as the region of interest. can also be extracted. Step 200 may be performed using a deep neural network-based convolutional neural network model trained with one or more learning data corresponding to a pre-selected region, and the region of interest is at least one of the frequency of fractures, the difficulty of detection, or the degree of sequelae when not detected. It may be an area selected based on one or more factors. If the target image is a wrist image, the region of interest may be a region corresponding to at least one of the radial bone, the steel bone, and the navicular bone.

Next, the electronic device may classify the abnormal state of the target image or the region of interest by using the machine learning model learned from the normal determination data and the abnormal determination data corresponding to the region of interest as learning data (S300). In this step, a classification value corresponding to the abnormal state of the target image or region of interest may be calculated. The classification value is a value calculated by applying the target image or region of interest to a deep neural network machine learning model, which is a layer included in the deep neural network. It may include at least one of an input value, an output value, or a calculation value calculated using the input value or the output value.

Next, the electronic device may segment the abnormal part in the target image or region of interest (S400). Segmentation means predicting to which class each pixel of a target image or region of interest belongs. That is, in this step, the electronic device may calculate an abnormal probability value for each pixel of the abnormal part. As shown in the figure, steps 300 and 400 may be performed concurrently or may be sequentially processed according to user settings, but are not particularly dependent on the preceding order.

The electronic device finally determines the abnormal state of the target image or region of interest by using the classification value corresponding to the abnormal state of the target image or region of interest calculated in step 300 and the abnormality 21 probability value for each pixel of the abnormal region calculated in step 400. It can be judged (S500).

More specifically, step 500 is described as follows.

In step 500, the electronic device may determine the abnormal state of the region of interest by applying the classification value and the abnormal probability value to a pre-learned integrated analysis model. It may be a neural network or a regression model learned using values and anomaly probability values for each pixel in the anomaly area as learning data.

Referring to FIG. 5 as another example of step 500, the electronic device counts the number of pixels for which the abnormality probability value exceeds a preset first threshold (S510), and the number of pixels in the target image or the region of interest After calculating the ratio (S530), and if any one of the calculated ratio and the classification value exceeds a preset second threshold, it can be determined that an abnormal part exists in the target image or the region of interest (S550). . Furthermore, the electronic device may adjust the first threshold according to the classification value. For example, when the classification value is a value corresponding to the presence of an abnormal region in the target image or the image of interest, the electronic device may automatically increase or decrease the first threshold to adjust the determination result, through which the electronic device It is possible to further increase the accuracy of the analysis of whether an abnormal part exists in the target image.

Referring to FIG. 6 as another example of step 500, the electronic device counts the number of pixels whose abnormal probability value exceeds a preset first threshold (S520), applies the abnormal probability value of the pixels to a preset function, and Calculation may be performed (S540). For example, various functions may be used according to user settings, such as calculating an average of abnormal probability values of pixels exceeding a threshold value or calculating a median value. The electronic device may determine an abnormal state of the target image or the region of interest using the calculated result value. For example, the electronic device may calculate an average of abnormal probability values of pixels exceeding a threshold value, and if the average value exceeds a preset third threshold value, it may be determined that an abnormal region exists in the target image or region of interest (S560). ).

Referring to FIG. 6 as another embodiment of step 500, the electronic device may determine whether to use the abnormal probability value according to the classification value (S501), and if it is determined to use the abnormal probability value, having a preset probability value An abnormal state of the region of interest may be determined using pixels (S503).

Meanwhile, referring to FIG. 4 again, the electronic device may further include step 600 of displaying a color or image corresponding to the abnormality probability value on the target image. In this case, as shown in FIG. area can be visualized.

In addition, in step 700, the electronic device may output at least one of the results calculated in steps 300 to 500, which is calculated by calculating the classification value calculated in step 300, the odds probability value for each pixel calculated in step 400, or the odds probability value. The representative value calculated in step 500 and a resultant value of whether the target image or region of interest includes an abnormal part may be output as a result.

The medical image analysis method according to the above-described embodiment may be implemented in the form of a recording medium including instructions executable by a computer, such as a program module executed by a computer. Computer readable media can be any available media that can be accessed by a computer and includes both volatile and nonvolatile media, removable and non-removable media. Also, computer readable media may include both computer storage media and communication media. Computer storage media includes both volatile and nonvolatile, removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules or other data. Communication media typically includes computer readable instructions, data structures, program modules, or other data in a modulated data signal such as a carrier wave, or other transport mechanism, and includes any information delivery media.

Some embodiments omitted in this specification are equally applicable when the subject of the implementation is the same. In addition, since the above-described present invention is capable of various substitutions, modifications, and changes without departing from the technical spirit of the present invention to those skilled in the art in the technical field to which the present invention belongs, the above-described embodiments and the appended It is not limited by drawings.

100: medical image analysis device
110: input unit
120: learning unit
130: region of interest extraction unit
150: anomaly classification unit
160: abnormal part detection unit
170: integrated analysis unit
190: output unit

Claims (14)

A method in which an analysis device analyzes a medical image,
step a of receiving a target image to detect an abnormal part;
b) extracting a region of interest from the target image;
c) classifying an abnormal state of the region of interest by using a machine learning model learned from the normal determination data and the abnormal determination data corresponding to the region of interest as learning data;
step d of segmenting the abnormal region in the ROI;
and an e step of determining an abnormal state of the target image or the region of interest using the classification value corresponding to the abnormal state calculated in step c and the abnormal probability value for each pixel of the abnormal region calculated in step d. Medical image analysis method.
According to claim 1,
The classification value is a value calculated by applying the region of interest to a deep neural network machine learning model, and is an input value or output value of a layer included in the deep neural network, or an operation value calculated using the input value or the output value. A medical image analysis method comprising at least one of
According to claim 1,
The step e is
determining an abnormal state of the region of interest by applying the classification value and the abnormal probability value to a pre-learned integrated analysis model;
The integrated analysis model is a neural network or regression model trained using classification values corresponding to abnormal states of a plurality of training images and abnormal probability values for each pixel in an abnormal region as training data.
According to claim 1,
The step e is
counting the number of pixels for which the oddity probability value exceeds a preset first threshold value;
calculating a ratio occupied by the number of pixels in the target image or the region of interest;
and determining that an abnormal region exists in the target image or the region of interest when either one of the calculated ratio and the classification value exceeds a preset second threshold.
According to claim 1,
The step e is
counting the number of pixels for which the oddity probability value exceeds a preset first threshold value;
and determining an abnormal state of the target image or the region of interest using a result value calculated by applying an abnormality probability value of the pixel to a predetermined function.
According to claim 4,
The medical image analysis method further comprising adjusting the first threshold according to the classification value.
According to claim 1,
The step e is
determining whether to utilize the abnormal probability value according to the classification value;
and determining an abnormal state of the ROI using a pixel having a predetermined probability value when it is determined that the abnormal probability value is used.
According to claim 1,
The area of interest is
A medical image analysis method that is a region selected based on at least one of fracture frequency, detection difficulty, or degree of sequelae when not detected.
According to claim 8
Extracting the region of interest
and extracting the region of interest using a convolutional neural network model based on a deep neural network learned with one or more training data corresponding to the previously selected region.
According to claim 1,
If the target image is a wrist image,
The region of interest is a region corresponding to at least one of a radius, a steel bone, and a navicular bone.
According to claim 1,
The medical image analysis method further comprising displaying a color or image corresponding to the abnormal probability value on the target image.
According to claim 1,
The medical image analysis method further comprising outputting at least one of the results calculated in steps c to e.
An apparatus for analyzing medical images,
an input unit that receives a target image to detect an abnormal part;
a region of interest extractor extracting a region of interest from the target image;
an abnormality classification unit that classifies whether or not there is an abnormal region in the region of interest by using a machine learning model learned from normal determination data and abnormality determination data corresponding to the region of interest as learning data;
an abnormal region detection unit configured to segment an abnormal region in the region of interest;
Medical treatment including an integrated analysis unit that determines the abnormal state of the region of interest by using the classification value corresponding to the presence or absence of the abnormal region calculated by the abnormal region classification unit and the abnormal probability value for each pixel of the abnormal region calculated by the abnormal region detection unit video analysis device.
A medical image analysis program stored in a computer readable medium to execute any one of the methods of claims 1 to 12.

Images