Artificial intelligence and medical image-based trigeminal neuralgia surgery planning navigation method

Abstract

A trigeminal neuralgia surgery planning navigation method based on artificial intelligence and medical images, comprising: I. obtaining a nerve medical image sequence and a blood vessel medical image sequence; II. if the two image sequences are image sequences of the same coordinate system at the same time, no registration is required; if the two image sequences are image sequences taken at different times or the image sequences taken at the same time are not in the same coordinate system, registration is required to place the two image sequences in the same coordinate system; III. using a deep neural network model to take the two image sequences in the same coordinate system as input, and outputting a trigeminal nerve image segmentation result and a blood vessel image segmentation result; IV. performing three-dimensional reconstruction on the trigeminal nerve image segmentation result and the blood vessel image segmentation result to obtain a three-dimensional model of the nerve blood vessel dissection result. The present application has fast processing speed and high reliability, and makes the preoperative assessment and surgery planning of the clinician simple, accurate and fast.

Description

TECHNICAL FIELD

[0001] The present application relates to the field of biomedical engineering and medical images, and particularly relates to a trigeminal neuralgia surgery planning navigation method based on artificial intelligence and medical images. BACKGROUND

[0002] Trigeminal neuralgia is a common clinical disease, which is mainly manifested as repeated paroxysmal severe pain in the trigeminal nerve distribution area of one side of the face. In the trigeminal nerve distribution area of the head and face, the disease suddenly occurs and stops, and the pain is lightning-like, knife-cutting-like, burning-like, stubborn, and unbearable. Speaking, washing face, brushing teeth or even walking in the wind can cause paroxysmal severe pain. The pain lasts for several seconds or minutes, and the pain is periodic.

[0003] The cause of trigeminal neuralgia is often due to vascular compression of the trigeminal nerve demyelination. The common methods for treating trigeminal neuralgia include the following: 1) radiofrequency treatment, the radiofrequency needle is punctured into the nerve and the nerve is damaged; 2) trigeminal microvascular decompression, in the operation, the trigeminal nerve and the responsible blood vessels compressing the nerve are separated and separated by a gasket; 3) balloon compression surgery, the trigeminal nerve half-month node is compressed to alleviate pain, but the surgery cannot solve the problem of the responsible blood vessels compressing the nerve; 4) gamma knife treatment, for elderly patients or patients who do not respond to surgery, the trigeminal nerve is damaged by high-energy radiation.

[0004] The above treatment methods, whether puncture, surgery or radiotherapy, need to be planned in detail in clinical treatment, especially the following points: 1) based on MR image, the compressed trigeminal nerve and the responsible blood vessels compressing the nerve are determined; especially for some patients, the nerve morphology, running variation is large, the responsible blood vessels can be a blood vessel, or can be a branch of arteriovenous, so it is particularly important to determine the shape and three-dimensional spatial relationship of the blood vessels and the nerve for treatment planning. 2) Microvascular decompression surgery is the most common primary trigeminal neuralgia surgery. For radiofrequency treatment, balloon compression surgery and other treatment methods, surgical treatment is also needed. In the surgical planning, the target point of the surgery needs to be determined, that is, the nerve damage position, the blood vessel and nerve separation position, or the nerve compression position. At the same time, due to the location of the trigeminal nerve near the brainstem, detailed surgical path planning is needed to prevent surgical accidents; for radiotherapy methods, the radiotherapy target point also needs to be determined to reduce the damage to other tissues, so it is particularly important to determine the three-dimensional spatial anatomical relationship of the trigeminal nerve and the blood vessels and find the target point.

[0005] Patent 202011336414.3 mentions a trigeminal neuralgia model based on MR images and its preparation method, which can meet the needs of clinicians for precise assessment and surgical planning of patients. However, this method has the following defects: 1) The cost of 3D printing is high, and the speed is slow, which makes it impossible to popularize in the clinic on a large scale; 2) The method mentions that the segmentation of the trigeminal nerve and blood vessels in the MR image needs to be manually outlined. Since the nerves and blood vessels in the MR image are often not in the same sequence, and the anatomical position is complex, manual segmentation requires high clinical skills and medical image recognition, and the segmentation speed is very slow, which cannot meet the clinical needs; 3) The 3D printed model can only be used for surgical teaching and surgical planning, and cannot be used for intraoperative navigation and positioning.

[0006] Patents 202110607517.7, 201910334939.4 and 202011141240.5 mention various image registration methods and segmentation methods for acoustic neuroma, liver tumors and skeletal structures based on artificial intelligence technology. However, the above methods are not suitable for trigeminal nerve surgery planning and navigation. First, the segmentation of the trigeminal nerve and the responsible blood vessels is much more difficult than acoustic neuroma, liver tumors and various skeletal structures due to their complex anatomical structure and complex anatomical relationship between blood vessels and nerves. Therefore, the design of the image segmentation neural network should have special features. Second, the target area of the trigeminal nerve and the related responsible blood vessels is very small, and specific adjustments should be made to the training method when training the neural network. Third, the magnetic resonance image sequence for observing the trigeminal nerve and blood vessels has special features, and not all magnetic resonance image sequences can be used as a data set for trigeminal nerve and blood vessel segmentation. The above patents do not mention specific neural image sequences, so the data set collection and annotation of the trigeminal nerve has special features. Finally, although patent 202110607517.7 also mentions a neural network for blood vessel segmentation and proposes to use the segmented blood vessels as a high-level area for surgery to avoid damaging blood vessels and causing bleeding during surgery, the intracranial blood vessels mentioned in patent 202110607517.7 are not consistent with the concept of responsible blood vessels described in this invention. In acoustic neuroma surgery, it is necessary to avoid damaging major intracranial arteries and veins, and the smaller branches of arteries and veins will not have a significant impact. However, the responsible blood vessels in trigeminal neuralgia are not major intracranial arteries and veins, but small arterial and venous branches. These branches are less visible in MRA images and more difficult to segment. Moreover, these small branch blood vessels are the main target of surgical operation in trigeminal neuralgia, but they are not the main high-risk blood vessels in other intracranial surgeries.

[0007] In summary, the three-dimensional image segmentation and reconstruction of the trigeminal nerve and its compression responsible blood vessels is time-consuming and difficult to promote in the clinic, which leads to difficulties in preoperative assessment and surgical planning for clinicians.

[0008] Therefore, how to solve the above-mentioned deficiencies existing in the prior art has become the subject to be studied and solved by the present application. SUMMARY

[0009] The application aims to provide a trigeminal neuralgia surgery planning navigation method based on artificial intelligence and medical images.

[0010] To achieve the above-mentioned purpose, the technical solution adopted by the present application is:

[0011] A trigeminal neuralgia surgery planning navigation method based on artificial intelligence and medical images, comprising:

[0012] Step one, obtaining a sequence of neural medical images and a sequence of vascular medical images;

[0013] Step two, confirming whether the two image sequences need to be registered;

[0014] If the two image sequences are image sequences of the same coordinate system taken at the same time, no image registration is needed;

[0015] If the two image sequences are image sequences taken at different times or the image sequences taken at the same time are not in the same coordinate system, image registration is needed, and the two image sequences are placed in the same coordinate system;

[0016] Step three, using a deep neural network model, taking the two image sequences in the same coordinate system as input, and outputting trigeminal nerve image segmentation results and vascular image segmentation results;

[0017] Step four, performing three-dimensional reconstruction on the trigeminal nerve image segmentation results and the vascular image segmentation results to obtain a three-dimensional model of the neural vascular dissection results.

[0018] The relevant contents in the above technical solution are explained as follows:

[0019] 1. In the above scheme, in step one, the sequence of neural medical images is the FIESTA sequence of general medical magnetic resonance or the CISS sequence of Siemens medical magnetic resonance; the sequence of vascular medical images is the MRA sequence of magnetic resonance or the CTA sequence of CT.

[0020] 2. In the above scheme, in step three, the deep neural network is two, one of which takes the sequence of neural medical images as input and outputs the trigeminal nerve image segmentation results, and the other of which takes the sequence of vascular medical images as input and outputs the vascular image segmentation results.

[0021] 3. The method of any of the preceding methods, wherein in step three, the deep neural network is one, the input is the sequence of neuro medical images and the sequence of vascular medical images, and the output is the trigeminal nerve image segmentation result and the vascular image segmentation result.

[0022] 4. The method of any of the preceding methods, wherein the deep neural network is a VNet3D neural network or a UNet3D neural network.

[0023] 5. The method of any of the preceding methods, wherein a neural network convolution module is added after the output layer of the deep neural network.

[0024] 6. The method of any of the preceding methods, wherein before step three, the deep neural network model is trained, the training process comprising: establishing an image segmentation data set, the image segmentation data set comprising data of a sequence of neuro medical images and a sequence of vascular medical images of no less than 50 patients; outlining and segmenting the trigeminal nerve region and the vascular region in the data to obtain a segmentation result; based on the data and the segmentation result, training the deep neural network model multiple times to obtain multiple deep neural network models; and selecting the deep neural network model with the highest image segmentation accuracy from the multiple deep neural network models obtained by training as the deep neural network model used in step three.

[0025] 7. The method of any of the preceding methods, wherein in the process of training the deep neural network model, a Generalized Dice Loss or a Generalized Wasserstein Dice Loss is used as the loss function for training the deep neural network model.

[0026] 8. The method of any of the preceding methods, wherein in step four, the algorithm for three-dimensional reconstruction is a Marching Cubes algorithm or a RayCasting algorithm.

[0027] The working principle and advantages of the present application are as follows:

[0028] The present application is a trigeminal neuralgia surgery planning and navigation method based on artificial intelligence and medical images, comprising: I. obtaining a sequence of neuro medical images and a sequence of vascular medical images; II. confirming whether the two image sequences need to be registered; if the two image sequences are image sequences taken at the same time in the same coordinate system, no registration is needed; if the two image sequences are image sequences taken at different times or image sequences taken at the same time but not in the same coordinate system, registration is needed to place the two image sequences in the same coordinate system; III. using a deep neural network model to input the two image sequences in the same coordinate system, and outputting a trigeminal nerve image segmentation result and a vascular image segmentation result; IV. performing three-dimensional reconstruction on the trigeminal nerve image segmentation result and the vascular image segmentation result to obtain a three-dimensional model of the neurovascular anatomy result.

[0029] Compared with the prior art, the present application provides a kind of full-automatic segmentation and three-dimensional reconstruction method of trigeminal nerve and responsibility blood vessel based on deep learning and medical image, which can assist clinical doctors to better visualize the anatomical structure, shape, compression degree of nerve and blood vessel before operation;Effectively carry out operation planning, and determine the target point of operation;While three-dimensional model is used as the input of operation navigation, it can assist clinical doctors to realize accurate operation, reduce operation complications, reduce unnecessary exploration process in operation, and speed up operation time.

[0030] The method of the present application has fast processing speed and high reliability, so that the preoperative assessment and operation planning of clinical doctors become simple, accurate and fast. BRIEF DESCRIPTION OF DRAWINGS

[0031] Figure 1 shows a flowchart of the embodiment of the present application. Figure 1

[0032] Figure 2 shows a nerve image diagram for identifying the trigeminal nerve region. Figure 2

[0033] Figure 3 shows a blood vessel image diagram for identifying the blood vessel region. Figure 3

[0034] Figure 4 shows a blood vessel nerve three-dimensional reconstruction result diagram for identifying the nerve and blood vessel compression position of the embodiment of the present application. Figure 4 DETAILED DESCRIPTION

[0035] The present application will be further described below in combination with the drawings and embodiments:

[0036] Embodiment: The present application will be clearly explained by figures and detailed description below. Any person skilled in the art can change and modify the technology taught by the present application without departing from the spirit and scope of the present application after understanding the embodiments of the present application.

[0037] The terms used herein are only for describing specific embodiments and are not intended to limit the present application. As for the "include", "include", "have" and the like used herein, they are all open terms, that is, they mean to include but not limited to.

[0038] As for the words (terms) used herein, except for special notes, they generally have the usual meaning of each word used in this field, in the content of the present application and in the special content. Some words used to describe the present application will be discussed below or elsewhere in this specification to provide additional guidance for those skilled in the art on the description of the present application.

[0039] Referring to the accompanying Figures 1 to 4 Figure 1 shows a kind of trigeminal neuralgia operation planning navigation method based on artificial intelligence and medical image, which includes:

[0040] ​​​​Step one, obtaining a neurological medical image sequence and a vascular medical image sequence;

[0041] The neurological medical image sequence can be a FIESTA sequence of a GE medical magnetic resonance or a CISS sequence of a Siemens medical magnetic resonance; and the vascular medical image sequence can be an MRA sequence of a magnetic resonance or a CTA sequence of a CT.

[0042] Step two, confirming whether the two image sequences need to be registered;

[0043] If the two image sequences are image sequences in the same coordinate system at the same time (for example, a patient uses a GE medical magnetic resonance to simultaneously shoot a FIESTA image sequence and an MRA image sequence), no image registration is needed;

[0044] If the two image sequences are image sequences shot at different times (for example, a FIESTA sequence is shot by a magnetic resonance first and a CTA sequence is shot by a CT examination later) or image sequences shot at the same time are not in the same coordinate system, image registration is needed to place the two image sequences in the same coordinate system.

[0045] The image registration method can be a mutual information image registration method or a deep neural network image registration method in deep learning, which is a conventional method and will not be described herein as it is not the point of the present application.

[0046] Step three, using a deep neural network model in deep learning to take the two image sequences in the same coordinate system as input, and output a trident nerve image segmentation result and a vascular image segmentation result;

[0047] The deep neural network can be two, one of which takes the neurological medical image sequence as input and outputs the trident nerve image segmentation result, and the other of which takes the vascular medical image sequence as input and outputs the vascular image segmentation result. Alternatively, the deep neural network can be one, which takes the neurological medical image sequence and the vascular medical image sequence as input and outputs the trident nerve image segmentation result and the vascular image segmentation result.

[0048] Preferably, the deep neural network is a VNet3D neural network or a UNet3D neural network, or an improved version of the two neural networks.

[0049] Preferably, a neural network convolution module such as PointRend can be added after the output layer of the deep neural network to further improve the accuracy of image segmentation.

[0050] Preferably, the deep neural network model can be trained before step three, and the training process includes:

[0051] 1. Establishing an image segmentation dataset, which includes data of neural medical image sequences and vascular medical image sequences of no less than 50 patients (preferably more than 450);

[0052] 2. Outlining and segmenting the trigeminal nerve region and the vascular region in the data to obtain a segmentation result; the outlining method is preferably manual outlining by an operator (such as a clinician or an image doctor), but is not limited thereto, and can also use ITK-SNAP software or 3D Slicer software for outlining;

[0053] 3. Training a deep neural network model based on the data and the segmentation result to obtain a plurality of deep neural network models;

[0054] 4. Selecting the deep neural network model with the highest image segmentation accuracy (Dice Ratio) from the plurality of deep neural network models obtained by training as the deep neural network model used in step three, and the image segmentation accuracy (Dice Ratio) of the model is greater than 0.9.

[0055] In the process of training the deep neural network model, since the trigeminal nerve and the responsible blood vessel region accounts for a very small proportion in the image, in order to reduce the precision shock in the training process and avoid the loss of training result precision, the loss function of the deep neural network model training should be specially adjusted, and in the process of calculating the loss function, the trigeminal nerve and the responsible blood vessel target region should be given a specific weight to increase the stability of the training and the precision of the neural network model obtained by training. Therefore, the generalized Dice Loss or Generalized Wasserstein Dice Loss can be preferably used as the loss function of the deep neural network model training.

[0056] Step four, performing three-dimensional reconstruction on the trigeminal nerve image segmentation result and the vascular image segmentation result obtained by step three to obtain a three-dimensional model of the neurovascular anatomy result (see Figure 4 ).

[0057] Preferably, the algorithm of the three-dimensional reconstruction can be the Marching Cubes algorithm of surface rendering or the Ray Casting algorithm of volume rendering, and can also be an improved type thereof.

[0058] The present application first proposes a complete medical image automatic segmentation and three-dimensional reconstruction scheme of the trigeminal nerve and the responsible blood vessels, based on which the anatomy of the trigeminal nerve and the responsible blood vessels compressed by the trigeminal nerve can be three-dimensionally visualized within a few seconds after the clinician inputs the medical images, better three-dimensionally visualizing the anatomical structure, shape and compression degree of the nerve and blood vessels; effectively planning the surgery and determining the surgical target point; at the same time, taking the three-dimensional model as the input of the surgical navigation, the precise surgery can be realized, the surgical complications can be reduced, the unnecessary exploration process in the surgery can be reduced, and the surgery time can be accelerated.

[0059] To verify the technical effect of the present application, the nerve image sequence and the blood vessel image sequence of 100 patients with trigeminal neuralgia were included, and the automatic three-dimensional reconstruction method of the present application and the manual three-dimensional reconstruction method of the clinician with neurosurgery and imaging experience were used to record the time of all patients, calculate the mean value, and compare the time efficiency and the accuracy of the method of the present application, and the comparison results are as follows:

[0060]

[0061] Table 1

[0062] The neural network of Table 1 runs on the NVIDIA 3090 GPU platform, and the time unit is second (s). As can be seen from Table 1, the neural network method of the present application can greatly improve the accuracy of reconstruction.

[0063] At the same time, the Dice Ratio of the trigeminal nerve segmentation result obtained by the method of the present application and the clinician segmentation result is as high as 95.8%, and the Dice Ratio of the blood vessel segmentation result and the clinician segmentation result is as high as 97.2%, which proves that the automatic segmentation of the present application and the clinician segmentation result have high consistency and can meet the clinical needs.

[0064] The above examples are only for illustrating the technical concept and characteristics of the present application, the purpose is to enable those skilled in the art to understand the content of the present application and to implement it, and it cannot limit the protection scope of the present application. Any equivalent changes or modifications made according to the spirit and essence of the present application shall be covered within the protection scope of the present application.

Claims

1. A trigeminal neuralgia surgery planning navigation method based on artificial intelligence and medical images, characterized in that: The application relates to a method for obtaining a three-dimensional model of a nerve-vessel dissection result. The method comprises the following steps: Step 1: obtaining a nerve medical image sequence and a blood vessel medical image sequence; Step 2: determining whether the two image sequences need to be registered; If the two image sequences are image sequences taken at the same time in the same coordinate system, image registration is not needed; If the two image sequences are image sequences taken at different times or the image sequences taken at the same time are not in the same coordinate system, image registration is needed, and the two image sequences are placed in the same coordinate system; Step 3: using a deep neural network model, taking the two image sequences in the same coordinate system as inputs, and outputting a trident nerve image segmentation result and a blood vessel image segmentation result; Step 4: performing three-dimensional reconstruction on the trident nerve image segmentation result and the blood vessel image segmentation result to obtain a three-dimensional model of the nerve-vessel dissection result.

2. The surgical planning and navigation method of claim 1, wherein: Before step 3, the deep neural network model is trained, and in the training process, specific weights are given to the trident nerve and the target area of the responsible blood vessel to increase the accuracy of the neural network model, and Generalized Dice Loss or Generalized Wasserstein Dice Loss is used as the loss function for training the deep neural network model. In step 1, the nerve medical image sequence is a FIESTA sequence of a general medical magnetic resonance or a CISS sequence of a Siemens medical magnetic resonance; 3. The surgical planning and navigation method of claim 1, wherein: The blood vessel medical image sequence is an MRA sequence of a magnetic resonance or a CTA sequence of a CT.

4. The surgical planning and navigation method of claim 1, wherein: In step 3, the deep neural network is two, one of which takes the nerve medical image sequence as input and outputs the trident nerve image segmentation result, and the other of which takes the blood vessel medical image sequence as input and outputs the blood vessel image segmentation result.

5. The surgical planning and navigation method of claim 1 or 3 or 4, wherein: In step 3, the deep neural network is one, which takes the nerve medical image sequence and the blood vessel medical image sequence as input and outputs the trident nerve image segmentation result and the blood vessel image segmentation result.

6. The surgical planning and navigation method of claim 5, wherein: The deep neural network is a VNet3D neural network or a UNet3D neural network.

7. The surgical planning and navigation method of claim 1, wherein: A neural network convolution module is added after the output layer of the deep neural network. Before step 3, the process of training the deep neural network model comprises the following steps: An image segmentation data set is established, which includes the data of the nerve medical image sequence and the blood vessel medical image sequence of not less than 50 patients; The trident nerve area and the blood vessel area in the data are outlined and segmented to obtain segmentation results; Based on the data and the segmentation results, the deep neural network model is trained multiple times to obtain multiple deep neural network models; 8. The surgical planning and navigation method of claim 1, wherein: Among the multiple deep neural network models obtained through training, the one with the highest image segmentation accuracy is selected as the deep neural network model used in step 3. In step 4, the algorithm for three-dimensional reconstruction is the Marching Cubes algorithm or the Ray Casting algorithm.

Images