Tissue segmentation method using medical image

Abstract

A tissue segmentation method using a medical image acquired by a medical device, according to one embodiment of the present invention, comprises the steps of: preparing a first medical image data set acquired from a first medical device; preparing a second medical image data set acquired from a plurality of second medical devices that are not the first medical device; combining the first medical image data set and the second medical image data set so as to form a training data set; using the formed training data set so as to train a tissue segmentation model; and using the trained tissue segmentation model so as to segment a tissue by using a medical image acquired from the first medical device, which is not included in the first medical image data set, and thus a method in which a data set combination and automatic configuration methodology is utilized to increase the performance of automatic segmentation of tissues and organs on the basis of a deep learning network for an image acquired from a specific device is provided.

Description

Tissue segmentation method using medical images

[0001] The present invention relates to a tissue segmentation method using medical images, and more particularly, to a tissue segmentation method using medical images that performs tissue segmentation on data acquired from a specific medical device using a tissue model learned using a dataset that combines data acquired from a specific medical device and data acquired from another medical device.

[0002] Medical image segmentation is essential for anatomical studies, identifying regions of interest such as tumors and lesions, and developing radiation treatment plans. [Non-patent Document 1] However, it is time-consuming and there is a shortage of medical imaging professionals to perform this task, leading to a growing demand for automated segmentation technology. [Non-patent Document 2] This demand, coupled with advances in computing power and deep learning technology, is currently leading to extensive research and development in the field of AI-based automatic segmentation algorithms. [Non-patent Documents 3, 4]

[0003] One of the challenges that must be considered during the training and application of AI-based automatic segmentation models is the diverse image characteristics of medical images. The gaps between 2D slices and differences in image resolution hinder model learning, and differences in image texture between medical devices limit the clinical application of these models. [Non-patent literature 5, 6]

[0004] One solution to this problem is data augmentation techniques that generate artificial synthetic data. [Non-patent document 7] However, this method has limitations in reflecting differences in image characteristics across medical device types. [Non-patent document 8]

[0005] Recently, nnU-Net, an automatic configuration methodology for effectively training data with diverse image features, has been studied. [Non-patent document 5] Based on this, TotalSegmentator derived a standardized configuration method from a generalized CT dataset of 1,204 individuals obtained using various acquisition settings. This approach enables effective learning from large-scale datasets and demonstrates applicability to various medical devices, demonstrating its potential for clinical use. [Non-patent document 9]

[0006] Despite these attempts, U-Net's convolutional method relies on estimating identical independent distributions, making it difficult to achieve stable segmentation performance for images of medical devices not included in generalized datasets. [Non-patent literature 10, 11]

[0007] Accordingly, there is a need to develop an efficient method that can achieve stable segmentation performance even for medical device images that are not included in generalized data sets.

[0008] The purpose of the present invention to solve the above problems is to provide a tissue segmentation method using medical images, which performs tissue segmentation on data acquired from a specific medical device using a tissue model learned using a dataset that combines data acquired from a specific medical device and data acquired from another medical device.

[0009] In order to solve the above problem, a method for segmenting tissues using medical images according to an embodiment of the present invention may include a step of preparing a first medical image dataset acquired from a first medical device; a step of preparing a second medical image dataset acquired from a plurality of second medical devices other than the first medical device; a step of combining the first medical image dataset and the second medical image dataset to form a training dataset; a step of training a tissue segmentation model using the configured training dataset; and a step of segmenting tissues using a medical image acquired from the first medical device that is not included in the first medical image dataset, using the trained tissue segmentation model.

[0010] Here, the medical image acquired from the medical device may be any one of CT (Computed Tomography), MR (Magnetic Resonance), and PET (Positron Emission Tomography).

[0011] Here, in the step of preparing the second medical image dataset, the second medical image dataset may be an image of the same modality as the first medical image dataset.

[0012] Here, in the step of preparing the second medical image dataset, the second medical image dataset may include data considering multiple pathologies, multiple acquisition protocols, and multiple medical institutions.

[0013] Here, in the step of preparing the second medical image dataset, data obtained from a medical device of the same manufacturer as the first medical device may be included in the second medical image dataset at a ratio of less than 10%.

[0014] Here, in the step of combining the first medical image dataset and the second medical image dataset to form a training dataset, the first medical image dataset may be less than 10% of the entire training dataset.

[0015] Here, the first medical image dataset may be 3% to 5% of the entire training dataset.

[0016] Here, in the step of training the tissue segmentation model using the above-configured training dataset, the training of the tissue segmentation model may be training from scratch without using a) transfer learning and b) fine-tuning methods.

[0017] Here, in the step of training the tissue segmentation model using the above-configured training dataset, the tissue segmentation model may have a U-net structure.

[0018] Here, the step of training the tissue segmentation model using the above-configured training dataset includes using a network model, wherein the network may include at least one of a 2d U-Net, a 3d U-Net, and a 3d U-Net cascade.

[0019] Here, the step of training the tissue segmentation model using the configured training dataset includes extracting features from the training dataset, and extracting features from the training dataset may include extracting pixel spacing, median shape, intensity property, and modality information.

[0020] Here, the step of training a tissue segmentation model using the above-configured training dataset includes extracting features from the training dataset and performing preprocessing based on the extracted features, and performing preprocessing based on the extracted features may include c) standardizing voxel spacing through resampling, d) intensity normalization, and e) automatic determination of a network structure based on the extracted features.

[0021] Here, in the step of training the tissue segmentation model using the configured training dataset, the number of training times (Epoch) may be adjusted according to the size of the configured training dataset.

[0022] Here, in the step of learning the tissue segmentation model using the above-configured training dataset, in the case of a modality other than CT, it may include f) additional preprocessing and g) manual adjustment of the loss function.

[0023] According to the present invention, by providing a method for segmenting tissues using medical images, there is provided an effect of providing a method for improving the performance of automatic segmentation of tissues and organs based on a deep learning network for images acquired from a specific device by utilizing a data set combination and automatic configuration methodology.

[0024] That is, a more accurate tissue segmentation method can be provided by using a method of performing tissue segmentation on data acquired from a specific medical device using a tissue model learned using a dataset that combines data acquired from a specific medical device and data acquired from another medical device.

[0025] Figure 1 is a flowchart for explaining a tissue segmentation method using medical images according to one embodiment of the present invention.

[0026] FIG. 2 is a flowchart showing the training process of a device-dependent dataset based model (DDSM) for automatic tissue and organ segmentation using a tissue segmentation method using medical images according to one embodiment of the present invention.

[0027] FIG. 3 is a flowchart showing the training and testing process of a device-dependent data set-based segmentation model (DDSM) and a generalized data-based segmentation model (GDSM) in a tissue segmentation method using medical images according to one embodiment of the present invention.

[0028] FIG. 4 is a diagram showing the overall results of comparing the prediction accuracy of a device-dependent data set-based segmentation model (DDSM) and a generalized data-based segmentation model (GDSM) in a tissue segmentation method using medical images according to one embodiment of the present invention.

[0029] FIG. 5 is a diagram showing the results of comparing the prediction accuracy of a device-dependent data set-based segmentation model (DDSM) and a generalized data-based segmentation model (GDSM) for each organ in a tissue segmentation method using medical images according to one embodiment of the present invention.

[0030] FIG. 6, FIG. 7 and FIG. 8 are diagrams showing the results of comparing the predicted results of a device-dependent data set-based segmentation model (DDSM) and a generalized data-based segmentation model (GDSM) with actual measurements in a tissue segmentation method using medical images according to one embodiment of the present invention.

[0031] FIG. 9 is a drawing showing a box-and-whisker plot in a tissue segmentation method using a medical image according to one embodiment of the present invention.

[0032] Hereinafter, with reference to the attached drawings, embodiments of the present invention will be described in detail so that those skilled in the art can easily implement the present invention. The present invention may be implemented in various different forms and is not limited to the embodiments described herein.

[0033] In order to clearly explain the present invention, parts that are not related to the description are omitted, and the same reference numerals are used for identical or similar components throughout the specification.

[0034] Throughout the specification, when a part is said to be "connected" to another part, this includes not only the cases where they are "directly connected" but also the cases where they are "electrically connected" with other elements intervening. Furthermore, when a part is said to "include" a component, this does not exclude other components, but rather includes other components, unless otherwise stated.

[0035] When a part is said to be "on" another part, it can be directly on top of the other part, or there can be other parts between them. Conversely, when a part is said to be "directly on" another part, there are no other parts between them.

[0036] The terms first, second, and third, etc., are used to describe, but are not limited to, various parts, components, regions, layers, and / or sections. These terms are used only to distinguish one part, component, region, layer, or section from another part, component, region, layer, or section. Accordingly, a first part, component, region, layer, or section described below may be referred to as a second part, component, region, layer, or section without departing from the scope of the present invention.

[0037] The terminology used herein is for the purpose of describing particular embodiments only and is not intended to limit the present invention. As used herein, the singular forms "singular" and "comprising" include plural forms as well, unless the context clearly dictates otherwise. The word "comprising" as used herein specifies a particular feature, region, integer, step, operation, element, and / or component, but does not exclude the presence or addition of other features, regions, integers, steps, operations, elements, and / or components.

[0038] Terms indicating relative space, such as "below" and "above," may be used to more easily describe the relationship of one part to another part depicted in the drawings. These terms are intended to encompass other meanings or operations of the device being used, along with the intended meaning in the drawings. For example, if a device in a drawing is turned over, some parts described as being "below" other parts will be described as being "above" the other parts. Therefore, the exemplary term "below" includes both the up and down directions. The device can be rotated 90 degrees or at other angles, and the relative space terms will be interpreted accordingly.

[0039] Although not defined otherwise, all terms, including technical and scientific terms, used herein have the same meaning as commonly understood by those of ordinary skill in the technical field to which the present invention pertains. Terms defined in commonly used dictionaries are further interpreted to have meanings consistent with the relevant technical literature and the present disclosure, and are not to be construed as ideal or overly formal unless otherwise defined.

[0040] Hereinafter, with reference to the attached drawings, embodiments of the present invention will be described in detail so that those skilled in the art can easily implement the present invention. However, the present invention may be implemented in various different forms and is not limited to the embodiments described herein.

[0041] FIG. 1 is a flowchart illustrating a tissue segmentation method using medical images according to an embodiment of the present invention. FIG. 2 is a flowchart showing a training process of a device-dependent dataset-based model (DDSM) for automatic tissue and organ segmentation using a tissue segmentation method using medical images according to an embodiment of the present invention. FIG. 3 is a flowchart showing a training and testing process of a device-dependent dataset-based segmentation model (DDSM) and a generalized data-based segmentation model (GDSM) in a tissue segmentation method using medical images according to an embodiment of the present invention. FIG. 4 is a diagram showing the results of an overall comparison of the prediction accuracies of a device-dependent dataset-based segmentation model (DDSM) and a generalized data-based segmentation model (GDSM) in a tissue segmentation method using medical images according to an embodiment of the present invention. FIG. 5 is a diagram showing the results of an organ-by-organ comparison of the prediction accuracies of a device-dependent dataset-based segmentation model (DDSM) and a generalized data-based segmentation model (GDSM) in a tissue segmentation method using medical images according to an embodiment of the present invention.

[0042] FIGS. 6, 7, and 8 are diagrams showing the results of comparing the predicted results of a device-dependent data set-based segmentation model (DDSM) and a generalized data-based segmentation model (GDSM) with actual measurements in a tissue segmentation method using medical images according to one embodiment of the present invention. FIG. 9 is a diagram showing a box-and-whisker plot in a tissue segmentation method using medical images according to one embodiment of the present invention.

[0043] Referring to FIGS. 1 to 9 together, a tissue segmentation method using a medical image according to an embodiment of the present invention may include a step (S100) of preparing a first medical image dataset acquired from a first medical device; a step (S200) of preparing a second medical image dataset acquired from a plurality of second medical devices other than the first medical device; a step (S300) of combining the first medical image dataset and the second medical image dataset to form a training dataset; a step (S400) of training a tissue segmentation model using the configured training dataset; and a step (S500) of segmenting a tissue using a medical image acquired from the first medical device that is not included in the first medical image dataset using the trained tissue segmentation model.

[0044] First, in a tissue segmentation method using a medical image according to one embodiment of the present invention, the medical image acquired from the medical device may be any one of CT (Computed Tomography), MR (Magnetic Resonance), and PET (Positron Emission Tomography).

[0045] The step (S100) of preparing the first medical image dataset may be, in a tissue segmentation method using a medical image acquired from the medical device, preparing the first medical image dataset acquired from the first medical device.

[0046] Meanwhile, the step (S500) of segmenting tissues using medical images acquired from the first medical device may be to segment tissues using the learned tissue segmentation model for medical images acquired from the first medical device that are not included in the first medical image dataset.

[0047] Ultimately, considering the step (S500) of segmenting tissues using medical images acquired from the first medical device, the step (S100) of preparing the first medical image dataset may be to prepare the first medical image dataset so as to be included in the training dataset in order to secure the accuracy of the step (S500) of segmenting tissues using medical images acquired from the first medical device.

[0048] The step (S200) of preparing the second medical image dataset may be to prepare the second medical image dataset using data acquired from a plurality of second medical devices, not the first medical device. That is, similarly to preparing the first medical image dataset to be included in the training dataset, data acquired from a plurality of second medical devices, not the first medical device, may be prepared as the second medical image dataset.

[0049] Meanwhile, in the step (S200) of preparing the second medical image dataset, the second medical image dataset may be an image of the same modality as the first medical image dataset.

[0050] Additionally, in the step (S200) of preparing the second medical image dataset, the second medical image dataset may include data considering multiple pathologies, multiple acquisition protocols, and multiple medical institutions.

[0051] Meanwhile, in the step (S200) of preparing the second medical image dataset, data obtained from a medical device of the same manufacturer as the first medical device may be included in the second medical image dataset at a ratio of less than 10%.

[0052] That is, the process of preparing the second medical image dataset may include other medical devices from the same manufacturer as the first medical device. In this case, data acquired from devices from the same manufacturer as the first medical device may be included at a rate of less than 10% of the entire second medical image dataset.

[0053] In addition, in the step (S300) of combining the first medical image dataset and the second medical image dataset to form a training dataset, the first medical image dataset may be less than 10% of the entire training dataset. In addition, more preferably, the first medical image dataset may be 3% to 5% of the entire training dataset.

[0054] That is, in the step (S300) of configuring the training dataset, the training dataset includes both the first medical image dataset and the second medical image dataset, but the proportion of the first medical image dataset is less than 10%, and conversely, the proportion of the second medical image dataset is more than 90%.

[0055] In particular, preferably, the proportion of the first medical image dataset may be 3% to 5% of the entire training dataset, and the remainder of the entire training dataset may include the second medical image dataset.

[0056] Meanwhile, in the step (S400) of training a tissue segmentation model using the above-described training dataset, the training of the tissue segmentation model may be done by training from scratch without using a) transfer learning and b) fine-tuning.

[0057] Typically, training deep learning models for medical image tissue segmentation requires a large amount of data. Therefore, large-scale general datasets collected from various medical devices (scanners) across multiple hospitals and institutions have been utilized. However, images from specific medical devices at specific hospitals may not be included in these general datasets, and their number may be limited. In such cases, conventional methods such as a) transfer learning or b) fine-tuning have been used.

[0058] However, in the step (S400) of training a tissue segmentation model using the configured training dataset among the tissue segmentation methods using medical images according to one embodiment of the present invention, the training of the tissue segmentation model means training from scratch without using a) transfer learning and b) fine-tuning methods.

[0059] In addition, in the step (S400) of training a tissue segmentation model using the configured training data set, the tissue segmentation model may have a U-net structure, and the step (S400) of training a tissue segmentation model using the configured training data set may include using a network model, wherein the network may include at least one of a 2d U-Net, a 3d U-Net, and a 3d U-Net cascade.

[0060] In addition, the step (S400) of training a tissue segmentation model using the above-described training dataset includes extracting features from the training dataset, and extracting features from the training dataset may include extracting pixel spacing, median shape, intensity property, and modality information.

[0061] Furthermore, the step (S400) of training a tissue segmentation model using the above-configured training dataset includes extracting features from the training dataset and performing preprocessing based on the extracted features, and performing preprocessing based on the extracted features may include c) standardizing voxel spacing through resampling, d) intensity normalization, and e) automatic determination of a network structure based on the extracted features.

[0062] Meanwhile, in the step (S400) of training the tissue segmentation model using the above-configured training dataset, the number of training times (Epoch) may be adjusted according to the size of the above-configured training dataset.

[0063] In addition, in the step (S400) of training a tissue segmentation model using the above-described training dataset, if the medical image is a modality other than CT (Computed Tomography), it may include f) additional preprocessing and g) manual adjustment of a loss function.

[0064] In medical imaging, modality generally refers to the method of image acquisition, such as CT, MR, and PET. Even for images of the same modality, image intensity and texture characteristics can vary depending on the medical device model used. Currently, convolutional neural networks (CNNs), which demonstrate outstanding performance in automatic medical image segmentation, are dependent on the training data set, and model performance deteriorates when data with characteristics not considered during training occurs. This suggests that to achieve automatic segmentation optimized for individual devices, data from the relevant medical device must be directly incorporated into the training process.

[0065] According to the present invention, a data combination method that utilizes both generalized data and single-device data enables superior model performance with a small amount of single-device data. This contributes to enhancing the usability of existing and publicly available data. Unlike fine-tuning, a type of domain adaptation, the automatic configuration methodology automatically optimizes hyperparameters and training settings without requiring manual user intervention. This enables effective learning of training data and contributes to improved user experience.

[0066] Hereinafter, with reference to the drawings, a method for segmenting tissue using medical images according to one embodiment of the present invention will be described in more detail.

[0067] FIG. 2 is a flowchart showing the training process of a device-dependent dataset-based model (DDSM) for automatic tissue and organ segmentation using a tissue segmentation method using medical images according to one embodiment of the present invention. FIG. 2 shows the entire pipeline of the training process of the device-dependent dataset-based model (DDSM). Referring to FIG. 2, the training process of the device-dependent dataset-based model consists of a device-dependent dataset organization step based on a data combination method, and data feature extraction, preprocessing, and model learning using nnU-Net, which is an automatic configuration methodology.

[0068] Data combination method refers to the method of adding single-device data to a generalized data set when composing the model's training data set.

[0069] The pipeline of nnU-Net, an automatic configuration methodology, consists of feature extraction, preprocessing and training settings, and learning.

[0070] Feature Extraction: Extract features from individual data sets that make up the input training data set and use them as information for preprocessing and training setup optimization. Extracted information includes pixel spacing, median shape, intensity property, and modality.

[0071] - Preprocessing: Training data is preprocessed to improve model learning efficiency. Resampling preprocessing is performed to standardize the voxel spacing of the entire data based on the pixel spacing and median size obtained during the feature extraction step. Furthermore, intensity normalization preprocessing is performed on the entire image based on pixel value distribution and modality information. For non-CT images, such as those obtained from electron microscopy, additional modality-specific preprocessing and manual adjustment of the loss function may be considered.

[0072] Training Settings: The network's detailed structure, including the number of convolutional layers and the pooling design for individual image dimensions, is determined using the inter-pixel spacing and median size. The batch and patch sizes used during training are optimized. Manual adjustment of the training cycle (epochs) may be considered depending on the size of the device-dependent dataset.

[0073] - Training: Model training is performed based on training settings determined based on extracted features. The training networks provided by nnU-Net are 2d U-Net, 3d U-Net, and 3d U-Net cascade. Model training is performed individually according to the user's selection. At this time, new networks or network configuration variations can be considered based on the characteristics of the segmented organ or image features.

[0074] To validate the proposed device-specific automatic segmentation model (DDSM) based on a combined data set, we trained a prior art generalized data set-based model (GDSM) and compared the performance of the two models. Both models were trained to segment 21 regions of thoracic and abdominal organs.

[0075] FIG. 3 is a flowchart illustrating the training and testing process of a device-dependent dataset-based segmentation model (DDSM) and a generalized dataset-based segmentation model (GDSM) in a tissue segmentation method using medical images according to an embodiment of the present invention. FIG. 3 illustrates the training and testing process of a generalized dataset-based segmentation model (GDSM) and a device-dependent dataset-based segmentation model (DDSM). The generalized dataset-based model (GDSM) used only 1,203 generalized data sets when constructing the training data set, and randomly segmented them at a training:validation = 95:5 ratio. The device-specific automatic segmentation model (DDSM) used 50 single-device data sets in addition to 1,203 generalized data sets, and randomly segmented them at the same ratio. The specific data composition according to this is as shown in Table 1.

[0076] Model#Training(General)#Training(Single)#Validation(General)#Validation(Single)#% of Total Patients Training / ValidationGDSM11420610120395 / 5DDSM114446594125395 / 5

[0077] FIG. 4 is a diagram showing the overall results of comparing the prediction accuracy of a device-dependent data set-based segmentation model (DDSM) and a generalized data-based segmentation model (GDSM) in a tissue segmentation method using medical images according to one embodiment of the present invention.

[0078] Figure 4 compares the performance of a generalized data set based model (GDSM), which is a conventional technology that does not use a combined data set method, and a device-specific automatic segmentation model (DDSM) that utilizes a combined data set proposed in the present invention, using three evaluation indices.

[0079] The average Dice similarity coefficient (DSC) is an indicator of the voxel-level similarity between the correct answer and the model's estimate. The closer its value is to 1, the better the performance. On the other hand, the Hausdorff distance (HD) and the asymmetric surface distance (ASSD) are indicators of the boundary distance error between the correct answer and the model's estimate. The closer its value is to 0, the better the model's performance. The specific figures for Fig. 4 are shown in Table 2. In Table 2, the mean and median values ​​of the three evaluation indices are shown, and better performance results are highlighted in bold and underlined.

[0080] Model DSC (mean) DSC (median) HD (mean) HD (median) ASSD (mean) ASSD (median) GDSM 0.9125 0.9343 11.996.604 0.9053 0.8784 DDSM 0.9246 0.93989.4935.132 0.583 10.7626

[0081] As can be seen in Fig. 4 and Table 2, the performance of the device-specific automatic segmentation model (DDSM) using dataset combination was higher than that of the generalized dataset-based model (GDSM). The average Dice similarity coefficient (DSC) of the device-specific automatic segmentation model (DDSM) was 0.9246, indicating greater pixel-to-pixel similarity than that of the generalized dataset-based model (GDSM). In addition, the average Hausdorff distance (HD) and the average asymmetric surface distance (ASSD) of the device-specific automatic segmentation model (DDSM) were 9.493 mm and 0.5831 mm, respectively, indicating smaller distance errors than those of the generalized dataset-based model (GDSM). When comparing the differences between models through metrics, the device-specific auto-segmentation model (DDSM) showed consistent improvements of 1.3%, 21%, and 36% in all metrics, respectively, based on the relative performance improvement rate (|DDSM-GDSM| / GDSM×100)) compared to the generalized dataset-based model. By adding only a small number (4%) of single-device data relative to the total dataset size, we were able to observe an increase in the model's performance for device-specific acquired images.

[0082] FIG. 5 is a diagram showing the results of comparing the prediction accuracy of a device-dependent data set-based segmentation model (DDSM) and a generalized data set-based segmentation model (GDSM) for each organ in a tissue segmentation method using medical images according to one embodiment of the present invention. FIG. 5 shows the values ​​of three transformation indices (DSCdiff, HDratio, ASSDratio) of the device-dependent automatic segmentation model (DDSM) and the generalized data set-based model (GDSM) for 21 individual organs. The transformation method of each indices is shown in Equations 1, 2, and 3 below.

[0083]

[0084]

[0085]

[0086] Here, G and D represent GDSM and DDSM, respectively, and the subscript i represents an individual organ.

[0087] As shown in Figure 5, the device-specific automatic segmentation model (DDSM) outperformed the generalized dataset-based model (GDSM) in most organs. The device-specific automatic segmentation model (DDSM) showed better DSCdiff and HDratio scores in 14 of the 21 organs, and better ASSDratio scores in 15 organs. Furthermore, the device-specific automatic segmentation model (DDSM) outperformed the generalized dataset-based model (GDSM) in all three mean values ​​of the deformation indices.

[0088] FIGS. 6, 7, and 8 are diagrams showing the results of comparing the predicted results of a device-dependent data set-based segmentation model (DDSM) and a generalized data set-based segmentation model (GDSM) with actual measurements in a tissue segmentation method using medical images according to an embodiment of the present invention. FIGS. 6, 7, and 8 are diagrams comparing the qualitative results between a device-dependent automatic segmentation model (DDSM) according to an embodiment of the present invention and a conventional generalized data set-based model (GDSM). Referring to FIGS. 6, 7, and 8, the generalized data set-based model (GDSM) showed a segmentation result with a disconnected shape in the large intestine (FIG. 6), and incomplete boundary errors (FIG. 7) and small hole errors (FIG. 8) were observed in the liver.

[0089] Fig. 9 is a diagram showing a box-and-whisker plot in a tissue segmentation method using a medical image according to one embodiment of the present invention. Fig. 9 is a box-and-whisker plot, and the horizontal lines of the box represent 25% (Q1, 1st quartile), 50% (Q2, median), and 75% (Q3, 3rd quartile) of the data, respectively. The whiskers at the top and bottom of the box (lines connected to the box) represent the maximum and minimum values ​​among the values ​​whose difference from Q3 and Q1 is within 1.5 times the interquartile range (IRQ = Q3 - Q1), and the dots represent outliers that fall outside this range.

[0090] Although the embodiments of the present invention have been described with reference to the attached drawings, those skilled in the art will understand that the present invention can be implemented in other specific forms without changing the technical spirit or essential characteristics thereof. For example, those skilled in the art may change the material, size, etc. of each component according to the application field, or may combine or substitute the disclosed embodiments to implement the present invention in a form not specifically disclosed in the embodiments of the present invention, but this also does not depart from the scope of the present invention. Therefore, the embodiments described above should be understood as illustrative and not limiting in all aspects, and such modified embodiments should be considered to be included in the technical spirit described in the claims of the present invention.

Claims

1. A method for segmenting tissue using medical images acquired from a medical device, A step of preparing a first medical image dataset obtained from a first medical device; A step of preparing a second medical image dataset obtained from a plurality of second medical devices other than the first medical device; A step of combining the first medical image dataset and the second medical image dataset to form a training dataset; A step of training a tissue segmentation model using the training data set configured above; and A step of segmenting a tissue using a medical image acquired from the first medical device that is not included in the first medical image dataset, using the learned tissue segmentation model; A method for segmenting tissue using medical images including:

2. In paragraph 1, The medical images obtained from the above medical devices are either CT (Computed Tomography), MR (Magnetic Resonance), or PET (Positron Emission Tomography). A method for segmenting tissue using medical images, characterized by:

3. In paragraph 1, In the step of preparing the second medical image dataset above, The above second medical image dataset is an image of the same modality as the above first medical image dataset. A method for segmenting tissue using medical images, characterized by:

4. In paragraph 1, In the step of preparing the second medical image dataset above, The above second medical image dataset includes data considering multiple pathologies, multiple acquisition protocols, and multiple medical institutions. A method for segmenting tissue using medical images, characterized by:

5. In paragraph 1, In the step of preparing the second medical image dataset above, Among the above second medical image data sets, data obtained from medical devices of the same manufacturer as the above first medical device are included at a rate of less than 10%. A method for segmenting tissue using medical images, characterized by:

6. In paragraph 1, In the step of combining the first medical image dataset and the second medical image dataset to form a training dataset, The above first medical image dataset is less than 10% of the entire training dataset. A method for segmenting tissue using medical images, characterized by:

7. In paragraph 6, The above first medical image dataset is 3% to 5% of the entire training dataset. A method for segmenting tissue using medical images, characterized by:

8. In paragraph 1, In the step of training the tissue segmentation model using the training dataset configured above, The learning of the above-mentioned tissue segmentation model is done by training from scratch, without using a) transfer learning and b) fine-tuning. A method for segmenting tissue using medical images, characterized by:

9. In paragraph 1, In the step of training the tissue segmentation model using the training dataset configured above, The above organizational division model has a U-net structure. A method for segmenting tissue using medical images, characterized by:

10. In paragraph 1, Using the training dataset configured above, the step of training the tissue segmentation model includes using a network model, The above network includes at least one of 2d U-Net, 3d U-Net and 3d U-Net cascade. A method for segmenting tissue using medical images, characterized by:

11. In paragraph 1, The step of training the tissue segmentation model using the training dataset configured above includes extracting features from the training dataset, Extracting features from the above training dataset includes extracting pixel spacing, median shape, intensity property, and modality information. A method for segmenting tissue using medical images, characterized by:

12. In paragraph 11, The step of training the tissue segmentation model using the training dataset configured above is It includes extracting features from the above training dataset and performing preprocessing based on the extracted features. Preprocessing based on the extracted features includes c) normalizing voxel spacing through resampling, d) intensity normalization, and e) automatic determination of network structure based on the extracted features. A method for segmenting tissue using medical images, characterized by:

13. In paragraph 1, In the step of training the tissue segmentation model using the training dataset configured above, Adjust the number of training times (Epochs) according to the size of the training dataset configured above. A method for segmenting tissue using medical images, characterized by:

14. In paragraph 1, In the step of training the tissue segmentation model using the training dataset configured above, For modalities other than CT, f) additional preprocessing and g) manual tuning of the loss function are included. A method for segmenting tissue using medical images, characterized by:

Images