Method for predicting bone strength on basis of deep learning

Abstract

The present invention relates to a method for predicting the bone strength of a subject by using a learning model trained with image data, bone mineral density data, and bone strength data of an animal bone. According to the present invention, the method for predicting bone strength on the basis of deep learning enables accurate bone strength to be quickly and conveniently predicted and, further, enables appropriate bone cement, bone graft material and the like to be recommended to a subject on the basis of the predicted bone strength.

Description

Deep Learning-Based Bone Strength Prediction Method

[0001] The present invention relates to a deep learning-based bone strength prediction method using medical images, and more specifically, to a deep learning-based bone strength prediction method including the steps of inputting analysis-use medical image data including bones of a subject and bone density data of the subject into an input unit, preprocessing the input analysis-use medical image data of the subject in a preprocessing unit, and transmitting the preprocessed analysis-use medical image data of the subject and the bone density data of the subject to an analysis unit to derive a bone strength value of the subject using a learning model learned with image data, bone density data, and bone strength data of an animal bone.

[0002]

[0003] Osteoporosis is defined as a skeletal disease characterized by impaired bone strength, leading to an increased risk of fracture. The prevalence of osteoporosis has been increasing due to the aging population and declining birth rates. This rise in prevalence is alarming because it also increases the risk of fracture. It is known that one in ten people with a femoral fracture, a common osteoporotic fracture, will die within a year, making fractures a serious public health problem. Bone strength is determined by 70% bone quantity and 30% bone quality. Bone quality is determined by bone turnover rate, structure, microdamage, and mineralization, but its measurement is currently difficult. Bone strength, however, relies on 70% bone quantity, or bone density. Therefore, bone density measurement has become a useful method for diagnosing osteoporosis, and the current definition of osteoporosis relies on bone density.

[0004] Bone mineral density (BMD) can be measured using the most common method, dual energy X-ray absorptiometry (DXA), at the midshaft of the spine, femur, and other extremities. The T-score of BMD measured at the midshaft using DXA serves as the WHO standard for diagnosing osteoporosis, and is only applied to postmenopausal women and men over 50 years of age. However, many fractures occur outside the T-score range of the osteoporosis diagnostic criteria, so even following this standard misses a large proportion of fractures. Another way to measure BMD is to look at bone tissue morphology for bone resorption parameters (eroded surface and number of osteoclasts). Bone tissue contains both living and dead osteoclasts, and this method can sometimes diagnose a low risk of osteoporosis because it can detect high bone density even when the dead osteoclasts, which cannot function properly, are accumulated in greater numbers than the living ones. In other words, despite the presence of a large number of dead osteoclasts, bone density values ​​are often overestimated, leading to an overlooked risk of fracture. For these and other reasons, the diagnosis of osteoporosis is increasingly focused on bone strength, not just bone density. Therefore, the development of new methods for measuring bone strength is urgently needed.

[0005] As we enter the era of big data and computing power increases, machine learning (ML) techniques are being widely applied in various medical fields, including bone diseases. Recent studies have actively utilized machine learning techniques to diagnose, detect, and predict fractures in osteoporosis, primarily using medical images. However, these attempts suffer from limitations such as the risk of overfitting when testing with trained data, decreased accuracy due to insufficient model validation, lack of representativeness of study subjects, lack of confidence intervals for statistical values, and unclear variable selection criteria. Consequently, osteoporosis risk prediction using machine learning techniques has not been widely adopted.

[0006] In addition, the existing technology for predicting bone density based on deep learning using medical images had the problem of being cumbersome and time-consuming because experts such as doctors had to directly select and input the target images used for bone density prediction from among several captured medical images. In addition, there was a problem that bone density prediction results for the same patient could vary depending on the selected images, making it difficult to obtain accurate bone density prediction results.

[0007] Moreover, as previously described, it is difficult to accurately predict fractures when diagnosis, detection, and fracture prediction of osteoporosis are performed solely based on bone density. Therefore, there is a need to predict osteoporosis risk from bone strength. However, most osteoporosis risk predictions using machine learning techniques have another limitation in that they are predicted only with data on bone density.

[0008] Accordingly, the inventors of the present invention have made extensive research efforts to overcome the problems of the above-mentioned prior arts, and as a result, have confirmed that when the bone strength value of the subject is derived through a learning model learned with image data, bone density data, and bone strength data of an animal's bones using analysis medical image data including the subject's bones and the subject's bone density data, accurate prediction of bone strength is possible even with a learning model learned with animal data, and thus have completed the present invention.

[0009] [Prior Art Document] KR10-2023-0007090 A

[0010]

[0011] Therefore, the main purpose of the present invention is to provide a deep learning-based bone strength prediction method capable of accurately predicting bone strength quickly and conveniently.

[0012]

[0013] According to one aspect of the present invention, a deep learning-based bone strength prediction method is provided, including the steps of inputting analysis-use medical image data including bones of a subject and bone density data of the subject into an input unit, preprocessing the input analysis-use medical image data of the subject in a preprocessing unit, and transmitting the preprocessed analysis-use medical image data of the subject and the bone density data of the subject to an analysis unit to derive a bone strength value of the subject using a learning model learned with image data, bone density data, and bone strength data of an animal bone.

[0014] In the deep learning-based bone strength prediction method of the present invention, the learned learning model may be trained by repeating the steps of obtaining a plurality of learning image data including bones of an animal and bone density and bone strength data for each of the learning image data, obtaining a plurality of preprocessed image data through preprocessing of the plurality of learning image data, and learning the correlation between the preprocessed image data and the bone density and bone strength of the bones of the animal corresponding to each preprocessed image data.

[0015] In the deep learning-based bone strength prediction method of the present invention, the preprocessing image data may be generated through a step of extracting a region of interest (ROI) from learning image data including animal bones, and a step of binarizing an image of the extracted region of interest (ROI).

[0016] In the deep learning-based bone strength prediction method of the present invention, the learned learning model may derive a correlation between the bone density and bone strength of the animal's bones by using a value obtained by counting pixels with a pixel value of 0 in a binarized image of the preprocessed image data.

[0017] In the deep learning-based bone strength prediction method of the present invention, the preprocessed medical image data for analysis of the subject may be generated through a step of extracting a region of interest (ROI) from learning image data including human bones, and a step of binarizing an image of the extracted region of interest (ROI).

[0018] In the deep learning-based bone strength prediction method of the present invention, the bone density data of the subject may be a T-score derived through dual-energy X-ray absorptiometry (DXA) analysis of the subject.

[0019] In the deep learning-based bone strength prediction method of the present invention, the bone density data of the animal's bone may be the animal's T-score derived through dual energy X-ray absorptiometry (DXA).

[0020] In the deep learning-based bone strength prediction method of the present invention, the animal may be at least one selected from the group consisting of pigs, cows, and horses.

[0021] In the deep learning-based bone strength prediction method of the present invention, the bone strength value can be derived as the human bone compressive strength (Mpa).

[0022] In the deep learning-based bone strength prediction method of the present invention, a step of recommending at least one of bone cement and bone graft material using the derived bone strength value through a deep learning-based algorithm may be further included.

[0023] According to another aspect of the present invention, the present invention provides a deep learning-based bone strength prediction system, including an input unit for inputting medical image data for analysis including bones of a subject and bone density data of the subject, a preprocessing unit for preprocessing the medical image data for analysis of the subject, an analysis unit for receiving the preprocessed medical image data for analysis of the subject and the bone density data of the subject and analyzing a bone strength value using a learning model learned with image data, bone density data, and bone strength data of an animal bone, and a result output unit for deriving the analyzed bone strength value.

[0024] In the deep learning-based bone strength prediction system of the present invention, the learning model of the analysis unit may be trained by repeating a step of learning the correlation between preprocessed image data of a plurality of learning image data including animal bones and a dataset including bone density and bone strength of the animal bones corresponding to each of the preprocessed image data.

[0025] In the deep learning-based bone strength prediction system of the present invention, the preprocessed image data may be generated by extracting a region of interest (ROI) from learning image data including animal bones, and then binarizing the image of the extracted region of interest (ROI).

[0026] In the deep learning-based bone strength prediction system of the present invention, the learned learning model may derive a correlation between the bone density and bone strength of the animal's bones by using a value obtained by counting pixels with a pixel value of 0 in a binarized image of the preprocessed image data.

[0027] In the deep learning-based bone strength prediction system of the present invention, the preprocessing data of the medical image of the subject for analysis input to the analysis unit may be generated through a step of extracting a region of interest (ROI) from learning image data including human bones and then binarizing the image of the extracted region of interest (ROI).

[0028] In the deep learning-based bone strength prediction system of the present invention, the bone density data of the subject input to the analysis unit may be a T-score derived through dual-energy X-ray absorptiometry (DXA) analysis of the subject.

[0029] In the deep learning-based bone strength prediction system of the present invention, bone density data of an animal's bones, which is one of the data for training the learning model of the analysis unit, may be the animal's T-score derived through dual energy X-ray absorptiometry (DXA).

[0030] In the deep learning-based bone strength prediction system of the present invention, the animal may be selected from the group consisting of pigs, cows, and horses.

[0031] In the deep learning-based bone strength prediction system of the present invention, the bone strength value derived through the result derivation unit may be derived as the human bone compressive strength (Mpa).

[0032] In the deep learning-based bone strength prediction system of the present invention, the result output unit of the system can recommend one or more of bone cement and bone graft materials through a deep learning-based algorithm using the bone strength value analyzed through the analysis unit.

[0033]

[0034] As described above, by using a learning model trained with image data, bone density data, and bone strength data for animal bones, accurate bone strength can be predicted quickly and conveniently, and further, appropriate bone cement, bone grafting material, etc. can be recommended to the subject based on the predicted bone strength.

[0035]

[0036] FIG. 1 is a diagram for explaining a deep learning-based bone strength prediction method according to an embodiment of the present invention.

[0037]

[0038] Hereinafter, the present invention will be described in more detail through examples and the attached drawings. These examples and drawings are intended solely to illustrate the present invention, and therefore, the scope of the present invention is not to be construed as being limited by these examples and drawings.

[0039]

[0040] FIG. 1 is a diagram for explaining a deep learning-based bone strength prediction method according to an embodiment of the present invention.

[0041] In an embodiment of the present invention, a deep learning-based bone strength prediction method first performs a step of inputting analysis medical image data including a subject's bones and the subject's bone density data into an input unit.

[0042] The medical imaging data including the bones of the subject may be one or more data selected from the group consisting of X-ray, computed tomography (CT), and dual energy absorptiometry (DXA), and preferably computed tomography (CT) data. The bone density data of the subject is a T-score derived through dual energy X-ray absorptiometry (DXA) analysis of the subject.

[0043] Medical image data including the bones of the subject input into the above input unit are generally preprocessed through an algorithm that binarizes the image. The medical image preprocessing data for subject analysis extracts a region of interest (ROI) from computed tomography (CT) image data including the bones of the subject, and binarizes the image of the extracted region of interest (ROI). More specifically, a region of interest (ROI) is extracted from CT image data including the bones of the subject to be analyzed. In order to increase the precision of the extracted analysis target image, the size is adjusted (resize), and in order to have a color with a normal distribution value, it is normalized (mormalized), and then the amount of data of the original image is increased (augmentation), and then the pixel values ​​of the analysis target image are compared with a predetermined threshold value to binarize so that the pixel values ​​are 0 or 255.

[0044] Next, a step is performed in which the subject's analysis data and the subject's bone density data, which have been preprocessed through the above steps, are transferred to an analysis unit and analyzed using a learning model trained on image data, bone density data, and bone strength data for the animal's bones.

[0045] The learning method of the above learning model first acquires a plurality of learning image data including animal bones and bone density and bone strength data for each of the learning image data.

[0046] The learning image data including the bones of the animal may be one or more data selected from the group consisting of X-ray, computed tomography (CT), and dual energy absorptiometry (DXA), and preferably computed tomography (CT) data. Such image data is composed of a plurality of pieces. In addition, the bone density data is the T-score of the animal derived through dual energy X-ray absorptiometry (DXA), and the bone strength is a value measuring the actual strength of the bones of each individual animal.

[0047] The plurality of training image data obtained above are preprocessed. The preprocessing of the plurality of training image data including animal bones is generally performed through an algorithm that binarizes the image. For example, the medical image preprocessing data for subject analysis extracts a region of interest (ROI) from computed tomography (CT) image data including the subject's bones, and binarizes the image of the extracted region of interest (ROI). More specifically, the region of interest (ROI) is extracted from CT image data including the bones of the animal to be analyzed. In order to increase the precision of the extracted analysis target image, the size is adjusted (resize), and in order to have a color distribution value, it is normalized (mormalized), and then the amount of data of the original image is increased (augmentation), and then the pixel values ​​of the analysis target image are compared with a predetermined threshold value to binarize the pixel values ​​so that they are 0 or 255.

[0048] As the final step of training the learning model, the correlation between the preprocessed image data for the animal bones and the bone density and bone strength of the animal bones corresponding to each of the preprocessed data is trained.

[0049] The above correlation is derived by using the value of pixels with a pixel value of 0 in the binarized image of the preprocessed image data to derive the relationship between the bone density and bone strength of the animal's bones. For example, in the case of the bone pore area in the binarized image data, since it has a pixel value of 0 (black), the more these 0 pixel values ​​there are, that is, the more pore areas there are, the more the estimated compressive strength can be estimated to be 'low' or 'low value', and from this, the relationship between the bone density and bone strength of the animal's bones can be derived.

[0050] Bones are composed of superficial bone and cancellous bone, of which superficial bone is observed through computed tomography (CT), and cancellous bone is observed through X-ray photography such as DEX. The learning model of the present invention learns the correlation between the T-score derived from cancellous bone and actual bone strength through DEX analysis of the animal and the compressive strength of superficial bone derived through preprocessing of CT image data including animal bones, that is, it learns the correlation by deriving features of the main components that make up the bone, so that more accurate bone strength values ​​for human bones can be predicted.

[0051] In particular, the animal may be one or more selected from the group consisting of a pig, a cow, and a horse, but preferably, by training with image data, bone density data, and bone strength data for pig bones, the same results can be derived when the learning results for animals are applied to humans.

[0052] Pigs are animals that are genetically close to primates, and in terms of genes that determine the shape of tissues and organs, humans and pigs are most similar at about 95%, and in particular, pig bones not only have mechanical properties (strength) and chemical composition similar to those of human bones, but also have a calcium to phosphorus ratio of 1.67 in pig bones, which is almost similar to that of human bones (1.68), so the bone strength prediction method of the present invention can be applied to humans by learning with image data, bone density data, and bone strength data for bones of animals, particularly pigs, and can derive accurate bone strength of humans.

[0053] Finally, the subject's bone strength value is derived from the subject's medical imaging data for analysis.

[0054] Bone strength analysis, i.e., deriving bone strength values ​​from the subject's medical image data for analysis, is performed using a deep learning-based artificial intelligence algorithm. While the AI ​​algorithm is not particularly limited, an ensemble model combining an Auto-Encoder (CNN) model capable of performing image processing on digital images and an LSTM model capable of processing the subject's information in text form is appropriate.

[0055] The above bone strength value is derived from the compressive strength of human bone (Mpa). The compressive strength of human bone varies depending on the person's age, gender, and body part, but generally, the compressive strength of human bone is 1.5 to 180 Mpa. Therefore, the bone strength value of the subject is derived from the compressive strength within the above range.

[0056] The deep learning-based bone strength prediction method of the present invention performs a step of selecting a recommendation list to provide useful information to a subject using the predicted bone strength results.

[0057] Specifically, a list of recommended products is selected, including one or more bone cements and bone grafting materials, and the selection of this list of recommended products is performed using a deep learning-based artificial intelligence algorithm.

[0058] The artificial intelligence algorithm for recommendation has information on various products similar to the subject's bone strength, and can select and recommend products with similar bone strength to the subject through the bone strength analysis results.

[0059] After the bone strength prediction as discussed above is completed, a result display step can be performed to visually display the analysis results by outputting them to a monitor or paper, etc., so that the subject can directly check these results or the examiner can use them as a reference when explaining the analysis results.

Claims

1. A step of inputting medical image data for analysis including the subject's bones and the subject's bone density data into an input unit; A step of preprocessing the medical image data for analysis of the subject input above in a preprocessing unit; and A deep learning-based bone strength prediction method comprising a step of transmitting the preprocessed medical image data of the subject for analysis and the bone density data of the subject to an analysis unit to derive the bone strength value of the subject using a learning model learned with the image data, bone density data, and bone strength data of the animal's bones.

2. In paragraph 1, The above learned learning model is, A step of acquiring a plurality of learning image data including animal bones and bone density and bone strength data for each of the learning image data; A step of obtaining a plurality of preprocessed image data through preprocessing of the plurality of learning image data; and A deep learning-based bone strength prediction method characterized in that it is trained by repeating the step of learning the correlation between the bone density and bone strength of the animal's bones corresponding to the above preprocessed image data and each preprocessed image data.

3. In paragraph 2, The above preprocessed image data is, A step of extracting a region of interest (ROI) from training image data containing animal bones; and A deep learning-based bone strength prediction method characterized in that it is generated through a step of binarizing an image of an extracted region of interest (ROI).

4. In paragraph 3, The above learned learning model is, A deep learning-based bone strength prediction method characterized in that a correlation between bone density and bone strength of the animal is derived by using a value obtained by counting pixels with a pixel value of 0 in a binarized image of the above preprocessed image data.

5. In paragraph 1, The above preprocessed medical image data for analysis of the subject is, A step of extracting a region of interest (ROI) from learning image data containing human bones; and A deep learning-based bone strength prediction method characterized in that it is generated through a step of binarizing an image of an extracted region of interest (ROI).

6. In paragraph 1, A deep learning-based bone strength prediction method, characterized in that the bone density data of the subject is a T-score derived through dual-energy X-ray absorptiometry (DXA) analysis of the subject.

7. In paragraph 1, A deep learning-based bone strength prediction method characterized in that the bone density data of an animal's bones is the animal's T-score derived through dual-energy X-ray absorptiometry (DXA).

8. In paragraph 1, A deep learning-based bone strength prediction method, wherein the animal is one selected from the group consisting of pigs, cows, and horses.

9. In paragraph 1, A deep learning-based bone strength prediction method, characterized in that the above bone strength value is derived as the human bone compressive strength (Mpa).

10. In paragraph 1, A deep learning-based bone strength prediction method, characterized in that it further includes a step of recommending at least one of bone cement and bone graft material using a deep learning-based algorithm using the derived bone strength value.

11. An input unit for inputting medical image data for analysis including the subject's bones and bone density data of the subject; A preprocessing unit that preprocesses the subject's medical image data for analysis; An analysis unit that receives preprocessed medical image data for analysis of a subject and bone density data of the subject and analyzes bone strength values ​​using a learning model trained with image data, bone density data, and bone strength data of animal bones; and A deep learning-based bone strength prediction system including a result output unit for deriving the analyzed bone strength value.

Images