KR20240110306A - Apparatus and method for identifying L5 spine in lumbar radiographic image - Google Patents

Abstract

The present invention relates to a device and method for identifying the L5 vertebra in lumbar radiography images. According to the present invention, in the device for identifying the L5 spine in a patient's lumbar spine radiology image, data collection is performed to collect training images in the form of a labeled L5 spine region in the lumbar spine radiograph image taken of the subject's lumbar spine for each of a plurality of subjects. wealth; A learning unit that learns the lumbar spine radiography image labeled with the L5 spine region and builds an artificial intelligence model to identify the L5 spine region from the lumbar spine radiology image; A data acquisition unit that acquires a lumbar radiography image of the lumbar region of a patient subject to analysis; and an area detection unit that applies the lumbar spine radiology image of the patient to be analyzed to the previously learned artificial intelligence model and provides the L5 spine area identified in the lumbar spine radiology image as a detection result.
According to the present invention, by inputting the patient's lumbar radiology image into a previously learned artificial intelligence model, the L5 spinal region can be recognized and provided within the lumbar radiographic image, and through this, clinicians can use the patient's L5 spinal region prior to spinal procedures or surgery. Allows you to quickly and accurately determine the location.

Description

Apparatus and method for identifying L5 spine in lumbar radiographic image}

The present invention relates to an apparatus and method for identifying the L5 vertebra in a lumbar radiology image, and more specifically, to a device for identifying the L5 vertebra in a lumbar radiography image that can identify the L5 vertebra by deep learning analysis of a patient's lumbar radiography image. It relates to devices and methods.

Accurate vertebral numbering in spine imaging is considered very important for pathological diagnosis and preoperative or preoperative planning. However, in cases where lumbosacral transitional vertebrae are present or when images of the entire spine are not available, vertebra numbering can be a difficult problem.

The lumbosacral metastatic spine includes the sacralization of lumbar vertebra 5 (hereafter L5) and the lumbosacralization of lumbar vertebra 1 (hereinafter S1). Sacralization of L5 is a congenital spinal deformity in which the long transverse process of the last lumbar vertebra is fused to varying degrees with the first sacral segment. Lumbarization of S1 is a spinal deformity in which the first and second parts of the sacrum are not fused, giving the appearance of six vertebrae in the lumbar spine.

The lumbar spine can be numbered accurately on a whole spine radiograph or whole spine MRI by counting down from the cervical 2 (C2) to the sacrum. If there is no image of the entire spine, numbering can be confusing.

It has been reported that approximately 30% of neurosurgeons have performed at least one incorrect spinal surgery. Therefore, clinicians need to photograph the patient's entire spine to accurately determine the height of the lumbar spine before performing spinal procedures or surgery.

Machine learning is a branch of artificial intelligence that automatically improves the ability to predict outcomes or diagnose pathology or disease through various trial-and-error approaches without explicit programming.

Deep learning is a technology that uses multiple layers of neural networks to process and analyze data, learn rules and patterns from given information, and extract meaningful information. Deep learning technology can surpass existing data analysis technologies and can learn from unstructured and perceptual data such as images and language and extract key feature information.

In particular, deep learning technology has excellent object detection capabilities. Object detection determines whether there are object instances of a specified category in an image and, if so, displays the location and extent of the object instances through a bounding box. Object detection, the foundation of image understanding and computer vision, serves as a foundation for solving complex or high-level vision tasks such as segmentation, object tracking, image captioning, activity recognition, and event detection. Object detection in the medical field is widely used to detect various pathologies such as malignancy, pneumonia, and fractures in medical images.

Therefore, a new technique is required that can quickly and accurately identify the L5 spine region through deep learning analysis of lumbar radiography images taken from patients.

The technology behind the present invention is disclosed in Korean Patent No. 10-2140393 (announced on July 31, 2020).

The purpose of the present invention is to provide a device and method for identifying the L5 spine in a lumbar spine radiology image, which can identify the L5 spine region in the lumbar spine radiology image by deep learning analysis of the patient's lumbar spine radiology image.

The present invention relates to a device for identifying the L5 spine in a patient's lumbar spine radiography image, wherein the data collection unit collects training images in the form of a labeled L5 spine region in the lumbar spine radiography image taken of the subject's lumbar spine for each of the plurality of subjects. ; A learning unit that learns the lumbar spine radiography image labeled with the L5 spine region and builds an artificial intelligence model to identify the L5 spine region from the lumbar spine radiology image; A data acquisition unit that acquires a lumbar radiography image of the lumbar region of a patient subject to analysis; and an area detection unit that applies the lumbar spine radiology image of the patient subject to analysis to the previously learned artificial intelligence model and provides the L5 spine area identified in the lumbar spine radiology image as a detection result.

Additionally, the training image may have a shape in which the L5 spinal region is labeled in the shape of a rectangular box in a radiographic image of the subject's lumbar spine.

Additionally, the artificial intelligence model for identifying the L5 spinal region can be implemented as the YOLOv5x model of the YOLOv5 product family.

Additionally, the data collection unit may collect training data through lumbar radiography images of a plurality of subjects, including patients with metal implants applied to the L5 spine and patients without metal implants.

Additionally, training data for patients to whom the metal implant is not applied may account for 5% to 10% of the total training data.

In addition, the present invention relates to a method of identifying the L5 spine performed by a device for identifying the L5 spine in a lumbar spine radiology image of a patient, wherein the L5 spine region is labeled in the lumbar spine radiograph image taken of the subject's lumbar spine for each of the plurality of subjects. collecting training images of shapes; Building an artificial intelligence model to identify the L5 spine region from the lumbar spine radiology image by learning a lumbar spine radiology image in which the L5 spine region is labeled; Obtaining a lumbar radiography image of the lumbar region of a patient to be analyzed; and applying the lumbar spine radiology image of the patient subject to analysis to the previously learned artificial intelligence model to provide the L5 spine region identified in the lumbar spine radiography image as a detection result.

According to the present invention, by inputting the patient's lumbar radiology image into a previously learned artificial intelligence model, the L5 spine region within the lumbar radiology image can be recognized and provided, and clinicians can determine the patient's L5 location prior to spinal procedures or surgery. Allows you to understand quickly and accurately.

Figure 1 is a diagram showing the configuration of a device for identifying the L5 spine in a lumbar spine radiography image according to an embodiment of the present invention.
FIG. 2 is a diagram briefly explaining the functions and data flow of each part of FIG. 1.
Figure 3 is a diagram illustrating details of a data set collected according to an embodiment of the present invention.
Figure 4 is a diagram illustrating a case in which the L5 spinal region is correctly detected from a lumbar spine radiography image in an embodiment of the present invention.
Figure 5 is a diagram illustrating a case in which the L5 spine region is incorrectly detected from a lumbar spine radiography image in an embodiment of the present invention.
FIG. 6 is a diagram illustrating a method of identifying the L5 spinal region using the device of FIG. 1.
Figure 7 is a diagram showing details of the deep learning algorithm used in an embodiment of the present invention.
Figure 8 is a diagram showing the performance of the L5 spine region detection model according to an embodiment of the present invention.

Then, 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 many different forms and is not limited to the embodiments described herein. In order to clearly explain the present invention in the drawings, parts unrelated to the description are omitted, and similar parts are given similar reference numerals throughout the specification.

Throughout the specification, when a part is said to be “connected” to another part, this includes not only cases where it is “directly connected,” but also cases where it is “electrically connected” with another element in between. . Additionally, when a part "includes" a certain component, this means that it may further include other components rather than excluding other components, unless specifically stated to the contrary.

FIG. 1 is a diagram showing the configuration of a device for identifying the L5 spine in a lumbar radiography image according to an embodiment of the present invention, and FIG. 2 is a diagram briefly explaining the functions and data flow of each part of FIG. 1.

Referring to Figures 1 and 2, the device 100 (hereinafter referred to as the identification device) for identifying the L5 spine in a lumbar spine radiography image according to an embodiment of the present invention includes a data collection unit 110, a learning unit 120, and a data collection unit 110. It includes an acquisition unit 130 and an area detection unit 140. The operation of each unit 110 to 130 and data flow between each unit may be controlled by a control unit (not shown).

The data collection unit 110 collects training images in which the L5 spine region is labeled within lumbar radiography images taken of the subject's lumbar region for each of the plurality of subjects. The collected data is used for machine learning. Here, the lumbar spine radiology image may mean an anteroposterior spine radiology image.

The data collection unit 110 may collect a training image in the form of a rectangular box labeling the L5 spine region in the anterior and posterior spine radiographic images for each subject, and provide the training image to the learning unit 120 .

The data collection unit 110 can build big data by collecting training images in which the L5 spinal region is labeled within the lumbar spine radiography image for each subject. This data collection unit 110 can collect basic data required for training through an EMR that manages and stores the patient's medical information or a hospital server linked thereto.

Data used for machine learning in the embodiment of the present invention was collected based on the following criteria. First, we retrospectively recruited 150 subjects aged 20 or older who visited our hospital from June 2016 to December 2021 and had anteroposterior spine radiographs taken. Then, the L5 vertebra on radiographs was determined by counting down from C2 to the sacrum. After creating images so that a rectangular box model represents the L5 spinal region, all annotation files were exported in YOLO format.

Figure 3 is a diagram illustrating details of a data set collected according to an embodiment of the present invention. As shown in FIG. 3, the data collection unit 110 trains through lumbar radiography images of a plurality of subjects, including patients who had metal implants (e.g., screws) applied to the L5 spine through surgery or the like and patients who did not have metal implants applied at all. Data can be collected.

Additionally, training data for patients who did not have metal implants applied can be set to account for 5% to 10% of the total training data, taking into account that there are more people who did not have surgery than people who had surgery. Here, surgery may mean inserting screws (metal implants) in the past.

The learning unit 120 learns an artificial intelligence model using the training images collected by the data collection unit 110. Specifically, the learning unit 120 builds an artificial intelligence model to identify the L5 spine region from the lumbar spine radiology image by learning the lumbar spine radiology image in which the L5 spine region is labeled.

In an embodiment of the present invention, the artificial intelligence model for identifying the L5 spinal region can be implemented as the YOLOv5x model of the YOLOv5 product family.

The learning unit 120 can implement an artificial intelligence model that identifies the L5 spine region from the spine radiology image by deep learning the spinal radiology image (training image) in which the subject's L5 spine region is labeled using the YOLOv5x algorithm.

Here, the learning unit 120 may be included or built into a processor as shown in FIG. 2. These processors are equipped with an L5 spinal region detection model using artificial intelligence and can provide final detection results. They can also be connected to external platforms, medical information systems, etc. to update or renew artificial intelligence learning data and detection models.

Once the artificial intelligence model is learned through learning, detection results of the L5 spinal region of the patient subject to analysis can be provided simply by inputting spine radiography images obtained from any patient subject to analysis into the artificial intelligence model.

To this end, the data acquisition unit 130 acquires a lumbar radiography image of the lumbar region of the patient being analyzed. The data acquisition unit 130 may search, query, and load lumbar radiography image data of the patient being analyzed in conjunction with the EMR or hospital server, system, medical staff terminal, etc., and acquire the data in real time from on-site X-ray equipment. You may.

Additionally, the data acquisition unit 130 can receive the patient's lumbar radiation image through a user terminal or various input means, and transmit the input data to the area detection unit 140.

Next, the region detection unit 140 applies the lumbar spine radiation image of the patient to be analyzed to the previously learned artificial intelligence model to identify the L5 spine region within the lumbar spine radiation image and provides the identified L5 spine region within the image as a detection result. do.

In other words, once the artificial intelligence model is completed through learning, the L5 spine region in the image can be identified and provided simply by inputting the lumbar radiation image observed for any patient subject to analysis into the artificial intelligence model.

The region detection unit 140 may provide the L5 spine region detection result for the lumbar spine radiation image of the patient to be analyzed, which is finally derived from the artificial intelligence model, to an output unit (not shown). Here, the output unit (not shown) can output and provide detection result data through a display, etc.

Figure 4 is a diagram illustrating a case in which the L5 spine region was correctly detected from a lumbar spine radiology image in an embodiment of the present invention, and Figure 5 illustrates a case in which the L5 spine region was incorrectly detected from a lumbar spine radiology image in an embodiment of the present invention. It is a drawing.

As shown in FIG. 4 , the area detection unit 140 may display and position the L5 identification area identified in the input lumbar spine radiography image in the form of a rectangular box. According to the algorithm of the embodiment of the present invention, the identification reliability of the L5 spinal region was more than 90%. However, looking at the incorrectly identified cases in FIG. 5, there are cases where the detection of the L5 spinal region itself fails, as shown in A and B of FIG. 5, and there are cases where inaccurate detection results are output, as shown in C.

The L5 spinal region identification device 100 according to an embodiment of the present invention may be a server itself that identifies the L5 spinal region from a spinal radiographic image, or may be a medical staff terminal device (PC, desktop, smartphone, laptop, pad, etc.) or an application implemented on a user terminal. Therefore, each terminal can be connected to the network with the identification device 100 while the related application (application program) is running and receive related services.

FIG. 6 is a diagram illustrating a method of identifying the L5 spinal region using the device of FIG. 1.

First, the data collection unit 110 collects training images in which the L5 spine region is labeled within lumbar radiographic images taken of the subject's lumbar region for each of the plurality of subjects (S610).

Next, the learning unit 120 builds an artificial intelligence model to identify the L5 spine region from the lumbar spine radiology image by learning the lumbar spine radiography image in which the L5 spine region is labeled (S620).

At this time, the learning unit 120 can learn an artificial intelligence model (L5 spine region detection model) for L5 spine region identification using YOLOv5x of the YOLOv5 family model. Additionally, the learning unit 120 can increase the reliability of the model through learning big data about related patients.

In the case of an embodiment of the present invention, YOLOv5x of the YOLOv5 family model was used to detect the L5 spine region, and YOLOv5 is a state-of-the-art object detection architecture family and corresponds to a model pre-trained on the COCO data set.

At this time, since the L5 spine region detection model according to the embodiment of the present invention detects only one class object, accuracy is more important than object detection speed, so the YOLOv5x model with the highest accuracy was used. In addition, as shown in Figure 3, among the images of 150 subjects, data from 105 (70%) were assigned to the training set, and data from 45 (30%) were assigned to the validation set.

Next, the data acquisition unit 130 acquires a lumbar radiation image of the lumbar region of the patient to be analyzed (S630). The data acquisition unit 130 may transmit the acquired data to the area detection unit 140.

Then, the region detection unit 140 applies the lumbar spine radiation image of the patient to be analyzed to the previously learned artificial intelligence model to identify the L5 spine region within the lumbar spine radiation image (S640), and the identified L5 spine region within the lumbar spine radiation image. is provided as a detection result (S650). Here, the L5 area detection result may be provided in the form shown in FIG. 4.

The following describes the machine learning algorithm used and the performance of the model of the present invention accordingly.

Figure 7 is a diagram showing details of the deep learning algorithm used in an embodiment of the present invention, and Figure 8 is a diagram showing the mAP (mean average precision) performance of the L5 spine region detection model according to an embodiment of the present invention.

The deep learning model used was implemented in Python 3.9.2, Pytorch 1.12.0, and scikit-learn, and the open source software LabelImg (Heartex, San Francisco, USA; https://github) was used to label the L5 spine. .com/heartexlabs/labelImg) was used. For details of the algorithm (details of hyperparameters), refer to FIG. 7.

The results of measuring the performance of the L5 detection model using mAP_0.5 and mAP_0.75 are shown in Figure 8. Here, the mAP_0.5 and mAP_0.75 metrics were considered “good” when there was Intersection over Union (IoU, an indicator of the degree of overlap between the area of the detected object and the area of the actual object), respectively. Specifically, mAP_0.5 was considered to have good performance when the IoU was 50%, and mAP_0.75 was considered to have good performance when the IoU was 75%. Figure 8 shows the change in mAP according to training epoch, and it can be seen that it shows a reliability of over 0.9 (90%).

As described above, according to the present invention, the L5 spine region can be identified within the lumbar spine radiology image by deep learning analysis of the patient's lumbar spine radiology image.

The present invention inputs the patient's lumbar spine radiography image into a previously learned artificial intelligence model (YOLOv5 model), recognizes and provides the L5 spine region within the image, and allows clinicians to identify the patient's spine area prior to spinal procedures or surgery. It allows you to quickly and accurately determine the location of L5.

The present invention has been described with reference to the embodiments shown in the drawings, but these are merely illustrative, and those skilled in the art will understand that various modifications and equivalent other embodiments are possible therefrom. Therefore, the true technical protection scope of the present invention should be determined by the technical spirit of the attached patent claims.

100: L5 spine identification device 110: data collection unit
120: Learning unit 130: Data acquisition unit
140: Area detection unit

Claims (10)

A device for identifying the L5 vertebra in a patient's lumbar spine radiograph, comprising:
A data collection unit that collects training images in the form of a labeled L5 spine region in a lumbar spine radiographic image taken of the subject's lumbar region for each of the plurality of subjects;
A learning unit that learns the lumbar spine radiography image labeled with the L5 spine region and builds an artificial intelligence model to identify the L5 spine region from the lumbar spine radiology image;
A data acquisition unit that acquires a lumbar radiography image of the lumbar region of a patient subject to analysis;
An L5 spine identification device comprising an area detection unit that applies the lumbar spine radiation image of the patient subject to analysis to the previously learned artificial intelligence model and provides the L5 spine region identified in the lumbar spine radiation image as a detection result. In claim 1,
The training image is an L5 spine identification device in which the L5 spine region is labeled in the shape of a rectangular box in a radiographic image of the subject's lumbar spine. In claim 1,
The artificial intelligence model for identifying the L5 spine region is an L5 spine identification device implemented as the YOLOv5x model of the YOLOv5 product family. In claim 1,
The data collection unit,
An L5 spine identification device that collects training data through lumbar radiography images of multiple subjects, including patients with metal implants applied to the L5 spine and patients without metal implants. In claim 4,
An L5 spine identification device in which training data for patients to whom the metal implant is not applied accounts for 5% to 10% of the total training data. A method of identifying the L5 vertebrae performed by a device for identifying the L5 vertebrae in a radiographic image of the lumbar spine of a patient, comprising:
Collecting training images in which the L5 spine region is labeled in lumbar radiography images taken of the lumbar region of each subject for each of the plurality of subjects;
Building an artificial intelligence model to identify the L5 spine region from the lumbar spine radiology image by learning a lumbar spine radiology image in which the L5 spine region is labeled;
Obtaining a lumbar radiography image of the lumbar region of a patient to be analyzed; and
An L5 spine identification method comprising applying the lumbar spine radiology image of the patient subject to analysis to the previously learned artificial intelligence model and providing the L5 spine region identified within the lumbar spine radiology image as a detection result. In claim 6,
The training image is a method of identifying the L5 spine in which the L5 spine region is labeled in the shape of a rectangular box in a radiographic image of the subject's lumbar spine. In claim 6,
The artificial intelligence model for identifying the L5 spine region is an L5 spine identification method implemented as the YOLOv5x model of the YOLOv5 product family. In claim 6,
The step of collecting the training data is,
An L5 spine identification method that collects training data through lumbar radiography images of multiple subjects, including patients with metal implants applied to the L5 spine and patients without metal implants. In claim 9,
An L5 spine identification method in which training data for patients to whom the metal implant is not applied accounts for 5% to 10% of the total training data.