Temporomandibular joint reposition post-evaluation method based on sagittal articular disc MRI image

Abstract

The invention discloses a temporal-mandibular joint reposition post-evaluation method based on a sagittal articular disc MRI (Magnetic Resonance Imaging) image. The method comprises the following steps: firstly, acquiring MRI images of an opening position and a closing position of a patient and information such as age, BMI and gender, performing denoising and enhancement preprocessing to improve image quality, and accurately segmenting the images by using CNN to determine joint disc and condylar process feature points; then multidimensional features such as the disc process angle, the articular disc thickness, the MRI signal intensity and the condylar process position are extracted in an all-around mode, and correlation coefficients among feature quantities are calculated; then, an improved LSTM model is adopted, joint changes in March, June and December after an operation are predicted according to the preoperative and intraoperative characteristic quantities, and the long-term influence of the operation is considered. And finally, integrating various data, inputting the integrated data into a multi-layer MLP, outputting a health overall score alpha2 (alpha2 belongs to [0.9, 0.8, 0.7, 0.6]), and realizing quantitative comprehensive evaluation. According to the method, the defects of a traditional method are overcome, joint information is accurately excavated, the dynamic prediction capacity is achieved, and a scientific and reliable basis is provided for evaluation and rehabilitation after temporomandibular joint reposition.

Description

Technical Field

[0001] The invention relates to the technical field of oral medical diagnosis, and in particular to a method for evaluating a temporomandibular joint after reduction based on sagittal articular disc MRI images. Background Art

[0002] The temporomandibular joint (TMJ) is located in front of the ears on both sides of the face, connecting the temporal bone and the mandible. It bears great pressure during chewing and other activities. Its articular disc acts as a cushion between the skull and the mandible, which is essential for the normal functioning of the joint. Temporomandibular joint disorder (TMD) is a complex disease caused by multiple factors, including malocclusion, mental stress, and fragile joint structure. It not only causes oral and facial pain, limited jaw movement, and other problems, but in severe cases it even affects maxillofacial development and chewing and speaking functions, reducing the patient's quality of life.

[0003] There are various surgical treatments for TMD at present, and accurate evaluation of the position of the disc after surgery is of great significance for the evaluation of surgical effects and patient rehabilitation. MRI is the gold standard for evaluating TMD disc displacement, which can accurately display the position of the disc and evaluate related lesions. However, although MRI-based quantitative analysis methods, such as the Drace standard (disc demarcation angle method) and the Hu standard (anterior disc demarcation method), are widely used in clinical practice, they have many shortcomings. The Drace standard relies on subjective interpretation and cannot reflect the dynamic changes of the disc; the Hu standard does not fully consider individual differences, and the pathological evaluation of disc morphology changes is not comprehensive. Therefore, a more complete postoperative evaluation method is urgently needed. Summary of the invention

[0004] To solve the above technical problems, the present invention provides a method for evaluating the temporomandibular joint after reduction based on sagittal articular disc MRI images, which is based on artificial intelligence, introduces an LSTM prediction model, and combines oblique sagittal closed-mouth and open-mouth MRI images to construct an articular disc.

[0005] The method for evaluating the temporomandibular joint after reduction based on sagittal articular disc MRI images of the present invention comprises the following steps: Data collection steps: Collect open-mouth MRI images, closed-mouth MRI images, and the patient's age, BMI, and gender information as basic data for subsequent evaluation; Image preprocessing step: De-noising and image enhancement processing are performed on the collected open and closed mouth MRI images to improve image quality and facilitate subsequent accurate segmentation and feature extraction; Image segmentation and feature point location steps: Use convolutional neural network (CNN) to segment the preprocessed MRI images of the articular disc and condyle, determine the locations of the condylar apex, condylar head center, and the midpoint of the posterior edge of the disc, providing a basis for calculating feature parameters; Evaluation steps of disc process angle and disc demarcation angle: define the coordinate system based on the orbitoauricular plane, determine the baseline for evaluating the disc process angle and disc demarcation angle according to the positions of the determined condyle and articular disc related points, and then calculate the disc process angle and disc demarcation angle. The disc process angle is used to judge the position status of the articular disc (such as favorable overcorrection, true overcorrection, occlusal stability, recurrence, etc.), and the disc demarcation angle is used to judge whether the articular disc has anterior displacement or lateral displacement; Disc thickness assessment steps: On the segmented disc image, measure the disc size in the x-direction (length) and y-direction (thickness), and calculate the correlation (correlation index) between the x-direction and y-direction changes to determine the disc shape migration and assess the disc status from the perspective of morphological changes; MRI signal intensity assessment steps: Quantify the grayscale value of the disc area in the MRI image to determine the signal intensity, analyze the mean value, standard deviation and other parameters of the signal intensity, set the brightness threshold to determine the effusion in the disc area, and assist in determining changes in disc tissue characteristics based on changes in signal intensity. Among them, the mean value represents the arithmetic mean of the signal intensity of all pixels in the articular disc area, reflecting the overall average level; the standard deviation measures the degree of dispersion of the signal intensity value relative to the mean value, reflecting the uniformity of the signal performance of the articular disc tissue on the MRI image; the maximum value determines the highest value that the signal intensity can reach in the articular disc area, which helps to detect areas with abnormally increased signal intensity; the minimum value reflects the lowest level of signal intensity in the articular disc area, assisting in the discovery of areas with abnormally low signal intensity; the median arranges the signal intensity values ​​in the articular disc area in ascending or descending order, and the value in the middle position (if the number of data is an odd number) or the average of the two middle values ​​(if the number of data is an even number) can more robustly reflect the data center trend; the skewness is used to describe the asymmetry of the signal intensity data distribution, helping to judge whether there is a lesion in the articular disc and its possible type from the perspective of the data distribution morphological characteristics; the kurtosis measures the steepness or flatness of the signal intensity data distribution relative to the normal distribution, and deeply understands the articular disc tissue in MRI The distribution morphological characteristics of signal intensity on the image; the signal intensity ratio corresponding to different tissue components, dividing the pixels in the articular disc area into different tissue component categories according to the preset threshold, calculating the proportion of each tissue component, reflecting the composition and changes of articular disc tissue components; the gradient change of signal intensity, by calculating the rate of change of signal intensity in the articular disc area in different directions, accurately locate the position and range of the lesion in the articular disc, and understand the degree of impact of the lesion on the surrounding tissue; quantiles (such as quartiles, deciles, etc.), quartiles sort the data from small to large and divide them into four equal parts of three values, and deciles divide the data into ten equal parts of nine values, which can describe the distribution of signal intensity data in more detail and discover potential abnormal distribution patterns; Condylar position assessment steps: measure the posterior condylar space (the distance between the condyle and the external auditory canal) and the anterior joint space, and use the anterior and posterior joint space base to calculate the joint space index, which accurately represents the displacement of the condyle relative to the central position and assesses the condylar position status; Model prediction steps: The improved LSTM model is used to analyze and predict the above-mentioned disc angle, disc demarcation angle, disc thickness, disc MRI signal intensity, and condylar position extracted before and during surgery, and the corresponding five feature quantities are obtained at 3 months, 6 months, and 12 months after surgery to consider the dynamic changes after surgery; Correlation calculation steps: Calculate the Pearson correlation coefficient, Spearman correlation coefficient, Kendall correlation coefficient and distance correlation analysis between the above five feature quantities to explore the internal connection between the feature quantities; Comprehensive evaluation steps: The original collected open and closed mouth MRI image data, the calculated feature quantities, the correlation coefficients between the feature quantities, and the patient's individual information such as age, BMI, and gender are integrated into a new data set and input into the multi-layer MLP. After being processed by the multi-layer MLP hidden layer, the compressed information variable is first output, and then the overall health score α2 (α2 belongs to [0.9, 0.8, 0.7, 0.6]) is further compressed to achieve a quantitative comprehensive evaluation of the postoperative state of the temporomandibular joint reduction, where 0.9 means that the joint recovery is excellent, close to or reaching the ideal postoperative state; 0.8 means that the recovery is good, but there are still slight observable problems; 0.7 means that the recovery is general, and there may be some situations that need attention or further intervention; 0.6 means that the recovery is poor, and the joint may still have many problems or be in an unstable state; the hidden layer of the multi-layer MLP is 6 layers, the first three layers have 50 nodes, the fourth and fifth layers have 10 nodes, the sixth layer has 5 nodes, and the output is 1 node. When the data is input into MLP, the original image data is converted into a format suitable for neural network input through a specific encoding method. For example, the image pixel values ​​are normalized and arranged in a certain order as part of the input vector. The feature quantity is directly used as the other part of the input vector and input together with the image data. The correlation coefficient is also integrated into the input vector after appropriate transformation (such as scaling or encoding). In the hidden layer, the nodes of each layer process the input data by weighted summation and perform nonlinear transformation through activation functions (such as ReLU functions). The design of 50 nodes in the first three layers helps to fully extract and integrate complex information in multi-source data. The 10 nodes in the fourth and fifth layers further abstract and compress the information. The 5 nodes in the sixth layer refine the information into more representative intermediate variables, and finally output the value of a node through the output layer, that is, the overall health score α2. This structural design helps to effectively integrate multi-source data information and accurately output comprehensive evaluation results.

[0006] Preferably, the input of the forget gate of the improved LSTM model is expanded to include the hidden states of the first two steps and the current input, and the update formula of the forget gate is: , where W f and b f are the weight and bias parameters of the forget gate, is the sigmoid activation function, is the hidden state at the previous time step (t-1), is the hidden state of the first two time steps (t-2), The input feature vector of the current time step (time t); This design aims to enhance the model's ability to process time series data and better characterize the impact of joint reduction on the position of the entire articular disc; Specifically, the hidden states h of the first two steps t - 1 、h t - 2 It carries important information about the position and shape of the articular disc in the previous time step, and is consistent with the input x in the current time step. t Working together on the forget gate, the model can comprehensively consider the changing trend of the articular disc state over a long period of time after joint reduction surgery when deciding whether to retain or forget information in the cell state. For example, during the recovery process of the articular disc after surgery, the changes in its position and morphology are often not completed instantly, but gradually evolve over multiple time steps. The introduction of the first two steps of hidden states allows the model to "remember" the articular disc state information from earlier times, thereby more accurately capturing this long-term dynamic change process and providing a more reliable basis for predicting the feature quantities at different time points after surgery.

[0007] Preferably, the improved LSTM model designs a new way to update the cell state, the formula is: , where W C and b C are the weights and bias parameters of candidate memory cells, It is a candidate memory cell, which is a candidate memory cell calculation method. In this way, the cell state update process is optimized to adapt to the data characteristics in the present invention.

[0008] Preferably, the cell state update formula of the improved LSTM model is: , where C t - 1 is the cell state at the previous time step, is a candidate memory cell, which is a candidate memory cell calculation method; ensuring the accuracy and effectiveness of cell state update, Represents the output of the forget gate, which is a gated signal with a value range of [0,1]. Its core function is to calculate the forget gate (usually based on the current input , previous hidden state, etc.), determines the proportion of information that needs to be retained in the cell state Ct-1 at the previous moment, the way to generate candidate memory cells, and updates the current cell state based on the output of the candidate memory cells and the forget gate. The two together constitute the cell state update mechanism of the improved LSTM model.

[0009] Preferably, the calculation of the output gate of the improved LSTM model remains unchanged. The output gate determines which information in the cell state will be output as a hidden state, and its formula is: ,in and b o It is the weight and bias parameter of the output gate; it reasonably outputs the hidden state information, calculates for the output gate, and determines which information in the cell state is output.

[0010] Preferably, the final hidden state of the improved LSTM model is determined by the output gate and the cell state, and the formula is: ; Ensure that the hidden state can accurately reflect the relevant information of the joint state. This formula is for hidden state calculation. The hidden state is the output gate ( ) acts on the cell state ( ) is used to pass the result to the next time step.

[0011] Preferably, the multi-layer MLP has 6 hidden layers, 50 nodes in the first three layers, 10 nodes in the fourth and fifth layers, 5 nodes in the sixth layer, and 1 node in the output; this structural design helps to effectively integrate multi-source data information and accurately output comprehensive evaluation results.

[0012] Preferably, when the disc process angle is between 0° and -60°, the position of the articular disc is favorable for overcorrection; when the disc process angle is greater than -60°, the position of the articular disc is truly overcorrected; when the disc process angle is between 0° and 90°, the articular disc is in a stable position dependent on the occlusion; when the disc process angle exceeds 90°, the position of the articular disc is in a recurrent state, providing a clear standard for judging the position status of the articular disc.

[0013] Preferably, anterior displacement is defined when the anterior-posterior angle between the tangent line of the posterior edge of the posterior band of the articular disc and the vertical line passing through the 12 o'clock direction of the condyle exceeds 10°; this is a basis for accurately defining the displacement of the articular disc.

[0014] Preferably, the calculation formula of the joint space index is: , where P represents the posterior joint space and A represents the anterior joint space; the change in the position of the condyle is accurately quantified by this formula, and the positive or negative of the numerator P−A directly reflects the size relationship between the posterior joint space and the anterior joint space. If α>0, it means P>A, and the condyle deviates backward; if α<0, it means P<A, and the condyle deviates forward; if α=0, then P=A, and the condyle is centered.

[0015] LSTM (Long Short-Term Memory) is a special type of recurrent neural network. LSTM aims to solve the problem of vanishing gradients encountered by traditional RNNs when processing long sequences and can effectively capture and process long-term dependencies. LSTM is mainly composed of a series of memory units, and each memory unit contains the following parts: Forget Gate, Input Gate, Candidate Memory Cell, Cell State, and Output Gate.

[0016] The beneficial effects of the present invention compared with the prior art are as follows: Accurate evaluation: By comprehensively considering multi-dimensional features such as the disc-condyle angle, the thickness of the articular disc, and the MRI signal intensity, and combining fine image preprocessing and accurate segmentation by CNN, it can comprehensively reflect the joint state, avoid one-sided evaluation, and provide sufficient basis for accurate judgment; Dynamic prediction: The improved LSTM model introduces the hidden states of the previous two steps, accurately captures the change of postoperative feature quantities over time, can prospectively predict the joint state at 3 months, 6 months, and 12 months after surgery, and give early warnings of problems, fitting the joint recovery trajectory; Scientific evaluation: The multi-layer MLP integrates the original data, feature quantities, correlation coefficients, and patient information, and outputs the quantitative score α2 through scientific hierarchical design. Combining with expert experience, it ensures objective and practical evaluation and helps doctors formulate strategies; Wide applicability: Multiple embodiments show that it is effective for patients of different ages, genders, medical histories, surgical types and data from different MRI devices, with strong versatility and can be stably applied in various medical institutions. Brief Description of the Drawings

[0017] Figure 1 is the data flow diagram of the present invention; Figure 2 is the image feature extraction flow chart of the present invention; Figure 3 is the schematic diagram of the reference line and the included angle position of the present invention; Figure 4 is the structural diagram of the improved LSTM of the present invention; Figure 5is a scoring flow chart of the present invention; Figure 6 is a diagram of the multi-layer MLP structure of the present invention; DETAILED DESCRIPTION

[0018] The specific implementation of the present invention is further described in detail below in conjunction with the accompanying drawings and examples. The following examples are used to illustrate the present invention, but are not intended to limit the scope of the present invention.

[0019] like Figures 1 to 6 As shown, The core technical solution of this invention covers the improvement of LSTM model construction and condylar position evaluation method. At the model level: Forget gate optimization: Input integration of the previous two steps of hidden state h t - 1 、h t - 2 With the current input x t ; Cell state update: first pass Calculate candidate memory cells and reuse them ( is the output of the forget gate) to complete the state update; Output and hidden state: The output gate passes through Calculate, the final hidden state is given by Sure.

[0020] At the condylar level, based on the joint space index formula (P is the posterior joint space, A is the anterior joint space) to quantify the condylar position. The following specific implementation method focuses on the above technical solution, and specifically presents the implementation details and application process of the technical solution through model training, parameter calculation, example verification and other links. The specific implementation process is as follows:

[0021] Example 1, (i) Data Collection

[0022] 1. A 35-year-old male patient with a BMI of 24 was selected to receive arthroscopic surgery for right temporomandibular joint disorder. The temporomandibular joint was scanned using a Siemens Magnetom Skyra 3.0T MRI scanner. The scanning parameters were set as follows: layer thickness 1.5mm, layer spacing 0.3mm, matrix size 256×256, field of view (FOV) 14cm, and high-quality MRI images of oblique sagittal closed and open mouth positions were obtained. At the same time, the patient's medical history was recorded in detail. The patient reported that the right temporomandibular joint had occasional pain when chewing in the past year, with no history of trauma, but had a long-term habit of grinding teeth at night.

[0023] 2. The scanned images are named in the format of "patient name_scan time_closed mouth position / open mouth position" and stored in a special medical imaging database. The patient's personal information and medical history are entered into the corresponding database table to ensure accurate association between the two. The radiologist will conduct a preliminary inspection of the image quality to confirm that there are no obvious motion artifacts in the image and the joint structure is clearly displayed.

[0024] 2. Image preprocessing 1. Use the non-local mean filtering algorithm to denoise the image. This algorithm calculates the similarity between the pixel points in the image and the pixels in its neighborhood, and performs weighted averaging on the pixel values, thereby removing noise while retaining image details. For the image in this example, the search window size is set to 21×21 pixels and the similarity window size is set to 7×7 pixels. After denoising, the image signal-to-noise ratio (SNR) is calculated. The SNR of the original image is 18.5dB, and after denoising, the SNR is increased to 25.2dB, and the peak signal-to-noise ratio (PSNR) is increased from 28.8dB to 33.5dB, indicating that the denoising effect is good.

[0025] 2. Use contrast-limited adaptive histogram equalization (CLAHE) method for image enhancement. Divide the image into small areas of 8×8 pixels, and set the contrast limit parameter to 40. After enhancement, subjective observation shows that the contrast of structures such as articular disc and condyle is significantly improved, and the boundaries are clearer; calculate the image entropy value. The entropy value of the original image is 6.2, and the entropy value after enhancement becomes 7.1, indicating that the amount of image information has increased and the enhancement effect is significant.

[0026] 3. Image segmentation and feature point positioning 1. Use the U-Net network structure based on deep learning to segment the image. 500 labeled MRI image samples of the temporomandibular joint were collected (350 for training, 100 for validation, and 50 for testing), which were jointly labeled by 3 experienced radiologists. The U-Net model was trained using the training set. The stochastic gradient descent (SGD) optimization algorithm was used during the training process. The initial learning rate was 0.001, the momentum parameter was 0.9, and the batch size was 16. After 300 rounds of training, the accuracy of the model on the validation set reached 92%. The image to be segmented was input into the trained model to obtain the segmentation results of the articular disc and condyle. The segmentation results were post-processed to remove isolated areas smaller than 50 pixels, fill holes, and obtain accurate segmentation contours.

[0027] 2. Determine the position of feature points on the segmented image. For the condylar apex, the point with the highest grayscale is found as the initial estimate by analyzing the grayscale value of the condylar region. The final condylar apex is determined in its 5×5 neighborhood according to the condylar shape characteristics, with coordinates (120, 80) (image coordinate system, with the origin at the upper left corner). The condylar head center point is determined by a region growing-based method. A point near the condylar region center is selected as the seed point. Region growing is performed according to the grayscale similarity criterion (grayscale difference threshold is 15), and the centroid coordinates of the growing region are calculated as the condylar head center point, with coordinates (130, 90). The midpoint of the posterior edge of the posterior disc band is fitted with a cubic spline curve to the contour of the articular disc, and the midpoint coordinates are calculated according to the geometric characteristics of the curve to be (100, 110). Two oral and maxillofacial surgeons independently checked and verified that the position of the feature points was accurate.

[0028] 4. Feature extraction 1. Define the coordinate system based on the orbitoauricular plane and calculate the disc process angle and disc demarcation angle. According to the determined coordinates of the characteristic points, the disc process angle is calculated using the vector angle formula, and the disc process angle is 30° (within the range of the occlusal stability position). Determine the tangent line of the posterior edge of the articular disc, and calculate its angle with the vertical line passing through the 12 o'clock direction of the condyle to be 8° (not exceeding 10°, no anterior displacement).

[0029] 2. Measure the thickness of the articular disc. Select 8 measurement points in the direction of the longest axis of the articular disc and measure its thickness in the y direction. The average value is 3.2 mm, and the length in the x direction is 15 mm. Calculate the correlation (correlation index) between the changes in the x and y directions. Using the Pearson correlation coefficient formula, the correlation coefficient is 0.72, indicating that the shape of the articular disc is relatively regular and the changes in the x and y directions have a strong linear relationship.

[0030] 3. Quantify the MRI signal intensity of the articular disc area, extract the grayscale value array of the pixel in the articular disc area, and calculate the average value to be 150, the standard deviation to be 30, the maximum value to be 220, the minimum value to be 80, the median to be 145, the skewness to be 0.3 (slightly positively skewed), and the kurtosis to be 2.8 (close to normal distribution). Set the brightness threshold to 200 (determined based on statistical analysis of a large number of samples), and find that the signal intensity of a small part of the posterior and lateral area of ​​the articular disc is higher than the threshold, indicating that there may be a small amount of effusion.

[0031] 4. Measure the posterior condylar space and the anterior joint space. The pixel distance of the posterior condylar space is 10 mm, and the pixel distance of the anterior joint space is 8 mm. According to the image resolution (each pixel represents 0.5 mm), the actual distance of the posterior condylar space after conversion is 5 mm, and the actual distance of the anterior joint space is 4 mm. Calculate the joint space index , indicating that the condyle is relatively moved posteriorly.

[0032] 5. Model prediction Build an improved LSTM model and set the forget gate weight and bias parameter W f , b f , adjust the cell state update related parameters W c , b c And the output gate parameter W o , b o . Initialize the hidden state h t-1 、h t-2 The model hyperparameters were determined, with the hidden layer set to 2 layers, the number of nodes in the first layer to 64, the number of nodes in the second layer to 32, the initial learning rate to 0.005, the number of training rounds to 300, and the Adam optimization algorithm used.

[0033] The five feature quantities extracted before and during surgery are combined into time series data, which are normalized and then input into the model. The model gradually calculates and updates the hidden state according to the improved calculation logic, and finally outputs the predicted values ​​of the five feature quantities 3 months, 6 months, and 12 months after surgery. The prediction results show that the disc-condylar angle is expected to become 25° 3 months after surgery, the thickness of the articular disc increases slightly to 3.3mm, the average MRI signal intensity becomes 140 (the effusion may decrease), and the posterior condylar gap becomes 4.5mm (the condylar position is improved); 6 months and 12 months after surgery, each feature quantity also shows a trend of gradual stabilization and improvement. By comparing with the actual partial review data 3 months after surgery (the actual review data is limited), the disc-condylar angle is actually 26°, and the prediction error is ,The prediction errors of other feature quantities are also within a reasonable range, indicating that the model prediction has a certain accuracy.

[0034] (VI) Correlation calculation The Pearson correlation coefficient, Spearman correlation coefficient, Kendall correlation coefficient and distance correlation analysis between the five characteristic quantities were calculated. For example, the Pearson correlation coefficient between the disc process angle and the thickness of the articular disc was 0.68 (which had a certain correlation with the previously calculated correlation index), indicating that there was a strong positive correlation between the two; the distance correlation coefficient between the articular disc MRI signal intensity and the effusion (judged by the signal intensity higher than the threshold area) was 0.8 (hypothetical value), further indicating that the signal intensity was closely related to the effusion.

[0035] The correlation coefficient is integrated with the original data and feature quantity into a new data set and input into the multi-layer MLP. The correlation coefficient is normalized in the MLP input layer to make it range between [0,1].

[0036] 7. Comprehensive evaluation Construct a multi-layer MLP with 6 hidden layers (50 nodes in the first three layers, 10 nodes in the fourth and fifth layers, 5 nodes in the sixth layer, and 1 node in the output). Determine the number of nodes in the input layer, and integrate the original image data (normalized and encoded), feature quantities, correlation coefficients, and patient age, BMI, gender, and other information into the input vector. For example, the image data is encoded as a 1000-dimensional vector (determined according to the image features and encoding method), with a total of 5 feature quantities, 4 correlation coefficients (each coefficient is represented by a node), and 3 patient personal information, and the total dimension of the input vector is 1012.

[0037] The input vector is fed into the MLP, and the data is calculated in the hidden layer. Each layer of nodes is processed by weighted summation and ReLU activation function. After multiple layers of calculation, the overall health score is finally output. , indicating that the patient recovered well after surgery, but there are still minor problems. Based on the scoring results, the doctor recommends that the patient continue rehabilitation training, such as mouth opening and closing exercises, and conduct regular checkups to observe the recovery of the joints.

[0038] Example 2, (i) Data Collection A 28-year-old female patient with a BMI of 21 underwent open surgery for left temporomandibular joint disorder. The GE Discovery MR750w 3.0T MRI scanner was used for scanning, with the following scanning parameters: layer thickness 1mm, layer spacing 0.2mm, matrix 256×256, FOV 12cm, and oblique sagittal closed and open mouth MRI images were obtained. The patient had a history of left temporomandibular joint pain for 2 years, a history of mild facial impact, and no habit of bruxism.

[0039] The image and information arrangement, storage and quality inspection are the same as in Example 1, and the image quality is good without obvious artifacts.

[0040] 2. Image preprocessing The bilateral filtering algorithm is used for denoising, which effectively removes noise while retaining edge information. The standard deviation in the spatial domain is set to 2, and the standard deviation in the grayscale domain is set to 0.1. After denoising, the SNR of the image is increased from 17.8dB to 24.5dB, and the PSNR is increased from 27.5dB to 32.8dB.

[0041] Adaptive gamma correction is used for image enhancement, and the gamma value is automatically adjusted according to the grayscale distribution of the image (in this example, the gamma value adjustment range is 0.6 - 1.2). After enhancement, the joint structure is more clear in subjective observation, and the calculated entropy value is increased from 5.8 to 6.8, with obvious enhancement effect.

[0042] 3. Image segmentation and feature point positioning The U-Net network structure is also used to segment the image, and the number of samples and the annotation method are the same as in Example 1. The training parameters are adjusted to a learning rate of 0.002, a momentum of 0.85, and a batch size of 32. After 400 rounds of training, the accuracy of the model in the validation set reaches 93%. The post-segmentation processing is the same as in Example 1, and an accurate segmentation result is obtained.

[0043] The method for positioning the feature points is the same as in Example 1, the coordinates of the condylar top are (115, 75), the coordinates of the condylar head center are (125, 85), and the coordinates of the midpoint of the posterior edge of the posterior disc are (95, 105). The doctor verified that the feature point positions are accurate.

[0044] 4. Feature extraction The calculated disc-condylar angle is -45° (in a true overcorrected state), and the disc demarcation angle: the angle between the tangent line of the posterior edge of the posterior band and the vertical line at the 12 o'clock direction of the condyle is 12° (there is anterior displacement).

[0045] The thickness of the articular disc was measured with an average value of 2.8 mm in the y direction and 13 mm in the x direction. The correlation coefficient between the x and y directions was 0.55 (the shape of the articular disc had some irregular changes).

[0046] MRI signal intensity quantification, mean 130, standard deviation 25, maximum 190, minimum 70, median 125, skewness 0.2 (slightly positive skewness), kurtosis 2.5 (close to normal distribution). The threshold was set to 180, and no obvious effusion area was found.

[0047] Condylar space measurement, posterior condylar space 8mm (actual 4mm), anterior joint space 10mm (actual 5mm), joint space index (Relative anterior displacement of the condyle).

[0048] 5. Model prediction The construction and parameter setting of the improved LSTM model are similar to those in Example 1, with 2 hidden layers, 64 and 32 nodes respectively, a learning rate of 0.006, and 350 training rounds.

[0049] The input feature quantity data was used to predict the postoperative feature values. The disc-condylar angle was expected to be -40°, the disc thickness was 2.9mm, the average MRI signal intensity was 125, and the condylar gap was 4.2mm 3 months after surgery. The feature quantities also showed corresponding change trends 6 and 12 months after surgery. Compared with the actual review data 3 months after surgery (partial feature quantity review), the disc-condylar angle was actually -38°, and the prediction error was about 5.3%, which was within the acceptable range.

[0050] (VI) Correlation calculation The correlation coefficients of characteristic quantities were calculated, such as the Pearson correlation coefficient between the disc process angle and the disc thickness was 0.62, and the correlation coefficient between the disc MRI signal intensity and the effusion distance was 0.1 (the effusion was not obvious and the correlation was low).

[0051] The relevant data are input into MLP and processed in the same way as in Example 1.

[0052] 7. Comprehensive evaluation The MLP construction and input data processing are the same as in Example 1, and the total dimension of the input vector is determined according to the actual data situation.

[0053] The final output is the overall health score , indicating that the patient's recovery is general. The doctor recommends strengthening rehabilitation training and closely observing joint symptoms. If any abnormality is found, seek medical attention in time.

[0054] Example 3, (i) Data Collection A 45-year-old male patient with a BMI of 26 underwent joint lavage for bilateral temporomandibular joint disorder. Philips Ingenia 3.0T MRI scanner was used for scanning, with scanning parameters: layer thickness 2mm, layer spacing 0.5mm, matrix 256×256, FOV 16cm, and oblique sagittal closed and open mouth images were obtained. The patient's medical history was bilateral joint pain with limited mouth opening for 6 months, a long-term habit of clenching teeth, and no history of trauma.

[0055] The image and information processing are the same as in Example 1, and the image quality is acceptable.

[0056] 2. Image preprocessing The wavelet threshold denoising method is used to set the threshold according to the coefficient characteristics of the image after wavelet decomposition. After denoising, the SNR of the image is increased from 16.5dB to 23.2dB, and the PSNR is increased from 26.2dB to 31.5dB.

[0057] Image enhancement is performed using histogram normalization, and the histogram distribution of the target image is adjusted based on the reference image. After enhancement, the joint structure is clear from subjective observation, and the entropy value is increased from 5.5 to 6.5.

[0058] 3. Image segmentation and feature point positioning The U-Net network was used for segmentation, and the sample conditions were the same as those in Example 1. The training parameters were learning rate 0.0015, momentum 0.92, batch size 24, 350 rounds of training, and the validation set accuracy was 91%. The post-segmentation processing was the same as that in Example 1.

[0059] Determine the position of the characteristic points, the coordinates of the condylar top (taking the right side as an example) are (122, 82), the coordinates of the condylar head center are (132, 92), and the coordinates of the posterior edge of the posterior disc are (102, 112). The doctor confirmed that it is accurate.

[0060] 4. Feature extraction The calculated disc-condylar angle (taking the right side as an example) is 15° (in the stable position relying on the occlusal splint), and the disc-division angle: the angle between the tangent line of the posterior edge of the posterior band and the vertical line at the 12 o'clock direction of the condyle is 5° (no anterior displacement).

[0061] The articular disc thickness was measured with an average value of 3.5 mm in the y direction and 16 mm in the x direction. The correlation coefficient between the x and y directions was 0.78 (good shape regularity).

[0062] MRI signal intensity quantification, mean 160, standard deviation 35, maximum 230, minimum 90, median 155, skewness 0.4 (positive skewness is obvious), kurtosis 3.0 (slightly deviated from normal distribution). The threshold was set to 210, and the signal intensity of a small area of ​​the anterior medial side of the articular disc was higher than the threshold, indicating the possibility of a small amount of effusion.

[0063] Condylar space measurement: posterior condylar space 12mm (actual 6mm), anterior joint space 9mm (actual 4.5mm), joint space index (Relative posterior displacement of the condyle).

[0064] 5. Model prediction The construction and parameter setting of the improved LSTM model are similar to those in Example 1, with 2 hidden layers, 64 and 32 nodes, a learning rate of 0.004, and 320 training rounds.

[0065] The input data predicts the postoperative feature quantities. At 3 months after surgery, the disc angle is expected to be 12°, the disc thickness becomes 3.6mm, the average MRI signal intensity drops to 150 (the effusion may be reduced), and the posterior condylar space becomes 5.5mm (the condylar position is fine-tuned); at 6 and 12 months after surgery, the feature quantities are continuously optimized and tend to be stable. By comparing the actual review data 3 months after surgery, the disc angle is actually 13°, and the prediction error is about 7.7%. The predicted values ​​of the other feature quantities are close to the actual review values, indicating that the model prediction has a certain reliability and can provide a reference for doctors to judge the development trend of the disease.

[0066] (VI) Correlation calculation A comprehensive analysis of the correlation between characteristic quantities showed that the Pearson correlation coefficient between the disc angle and disc thickness reached 0.7, indicating that the two were closely related; the distance correlation coefficient between the disc MRI signal intensity and the effusion was 0.75, further proving the effectiveness of signal intensity in judging effusion. At the same time, the condylar position and disc thickness also showed a certain correlation, with a correlation coefficient of 0.4, which means that condylar displacement may affect the morphological changes of the disc.

[0067] The correlation coefficient is integrated with the original data and feature quantities, and then input into the multi-layer MLP after normalization. For example, for the correlation coefficient in this case, a specific scaling algorithm is used to make it match the other input data dimensions to ensure that the MLP can make full use of this information to mine potential patterns.

[0068] 7. Comprehensive evaluation The model is built according to the established multi-layer MLP architecture, and the number of nodes in each layer is accurately set according to the input data dimension. In this example, the original image data is converted into a 1200-dimensional vector through complex encoding, combined with 5 feature quantities, 4 correlation coefficients and 3 personal information of the patient, to form an input vector with a total dimension of 1212.

[0069] The input vector is operated layer by layer by MLP. The first three layers of 50 nodes fully integrate the complex features of multi-source data, the middle two layers of 10 nodes efficiently compress information, and the last layer of 5 nodes extracts key intermediate variables, and finally outputs the overall health score α2 = 0.8. This score shows that the patient has a good recovery after surgery. Based on the score and detailed feature data, the doctor recommends that the patient maintain moderate rehabilitation training, regular follow-up monitoring, and prevent the risk of recurrence.

[0070] Comparative Example 1 1. Data collection and processing A 35-year-old male patient in Example 1 with a BMI of 24 was selected. He had a long-term habit of nighttime bruxism and right temporomandibular joint disorder and received arthroscopic surgery. The same Siemens Magnetom Skyra 3.0T MRI scanner was used, but the standard procedure was not strictly followed during the scan, resulting in slight motion artifacts in the image. After collecting the medical history information, the information was not systematically organized and deeply analyzed, but simply recorded.

[0071] The image was only subjected to conventional mean filtering for denoising, and the filter kernel size was selected based on experience without fully considering the image noise characteristics. Although some noise was removed, the edge details of structures such as the articular disc and condyle were blurred. Image enhancement processing was not implemented, making the image with artifacts and poor contrast more difficult to analyze in the subsequent analysis, which formed a sharp contrast with the high-quality image after fine processing in Example 1.

[0072] 2. Feature extraction The traditional manual segmentation method is used to outline the articular disc and condyle. Although the operator has some experience, the segmentation boundary is not accurate due to image quality and subjective factors, and deviations occur in many places. When determining the feature points, there is a lack of accurate positioning algorithms. For example, the top of the condyle is determined by rough observation, with an error range of ±3 pixels; the center point of the condyle head relies on simple estimation, with a deviation of about ±2 pixels; the midpoint of the posterior edge of the posterior disc band is not accurately positioned, and the front, back, left and right deviations can reach ±4 pixels. Based on these inaccurate segmentation and feature point positioning, the calculated disc angle error is as high as ±10°, which cannot accurately reflect the true position of the articular disc.

[0073] The measurement of disc thickness is quite arbitrary, and only a few positions are roughly measured without considering the correlation between the x and y directions. The obtained thickness value lacks representativeness and cannot fully reflect the morphological changes of the disc. The MRI signal intensity analysis is extremely simple, and only the grayscale of the image is roughly observed. The key parameters such as the mean value and standard deviation are not quantified. Whether there is effusion in the disc area is judged only by the naked eye, which has no scientific basis and is very different from the comprehensive and detailed feature extraction in Example 1.

[0074] 3. Evaluation Method The improved LSTM model in Example 1 was not used for postoperative prediction. Instead, a broad estimate of postoperative recovery was made based simply on preoperative characteristics and the doctor's limited clinical experience. For example, vague conclusions such as "there should be some improvement after surgery" and "there may not be major problems" were made. The dynamic change trend of feature quantities such as the position and morphology of the articular disc after surgery was completely ignored, which lacked foresight.

[0075] In the comprehensive evaluation stage, no multi-layer MLP was constructed for data fusion and analysis. The scattered information collected was simply listed, and the doctor directly gave a qualitative evaluation based on his subjective experience, such as "the recovery is general". It was impossible to output an accurate and quantitative overall health score like in Example 1, and it was difficult to measure the patient's actual recovery degree after surgery, nor could it provide effective guidance for subsequent precision treatment.

[0076] 4. Results comparison A follow-up examination 3 months after surgery revealed that abnormal changes in disc thickness caused mild pain, which was not detected in early measurements; tissue changes reflected by MRI signal intensity caused occasional joint jams, which had been missed in previous judgments. Due to inaccurate early assessments and failure to adjust the rehabilitation plan in a timely manner, the patient's recovery progress was slow and symptoms continued to plague daily life, in stark contrast to the orderly recovery of the patient in Example 1 based on accurate assessments.

[0077] Compared with Example 1, under the traditional treatment method of this comparative example, the patient missed the best time for rehabilitation due to the lack of accurate preoperative evaluation and postoperative prediction, and the joint function recovery was poor, which fully highlights the outstanding value of the present invention in accurate evaluation and effective intervention.

[0078] Comparative Example 2 1. Data collection and processing A 28-year-old female patient in Example 2 was selected, with a BMI of 21, who underwent open surgery for left temporomandibular joint disorder and had a history of mild facial impact. The same GE Discovery MR750w 3.0T MRI scanner was used for scanning, but the scanning parameters were improperly set, with a layer thickness of 1.5 mm (originally 1 mm) and a layer spacing of 0.5 mm (originally 0.2 mm), resulting in decreased image resolution and blurred display of some structures. No correlation analysis was performed after the medical history was collected, and the potential injury pattern of the articular disc and condyle caused by the history of facial impact was not considered.

[0079] For image denoising, median filtering was used, but the window size was not selected appropriately, which failed to effectively remove noise and destroyed the texture information of some images. Image enhancement used histogram equalization with uniform parameters. Excessive enhancement caused some areas of the image to lose their original layering, and the edges of the articular disc and condyle appeared too bright or too dark, which was not conducive to subsequent accurate segmentation. This was far from the image preprocessing effect that was finely optimized in Example 2.

[0080] 2. Feature extraction Semi-automatic segmentation software was used to assist in segmenting the articular disc and condyle. Although it was an improvement over manual segmentation, the software algorithm was old and sensitive to image blur and noise. The segmentation results were broken and discontinuous in many places, making them difficult to repair. When determining the feature points, due to inaccurate segmentation, the deviation of the condyle top was ±4 pixels, the deviation of the condyle head center was ±3 pixels, and the deviation of the midpoint of the posterior disc was ±5 pixels, which led to a deviation of about ±12° in the calculation of the disc angle. It could not accurately reflect the position relationship of the articular disc, which was much inferior to the accurate feature extraction in Example 2.

[0081] The measurement of disc thickness is affected by segmentation, the measurement path is not selected reasonably, the measured value fluctuates greatly, and the correlation between the x and y directions is not deeply analyzed, so the morphological change trend of the disc cannot be accurately judged. The MRI signal intensity statistics are biased, the interference of fuzzy areas is not excluded, and the effusion is misjudged, which is inconsistent with the actual pathological state of the joint, which is in contrast to the rigorous signal intensity analysis in Example 2.

[0082] 3. Evaluation Method The traditional linear regression model is used to predict the postoperative situation. This model does not take into account the nonlinear characteristics and complex changes of the time series in the recovery process of the temporomandibular joint. It only establishes a simple relationship based on a limited number of preoperative features, and the prediction results are far from the actual results. For example, the prediction of the change trend of the disc thickness 3 months after surgery is completely wrong and cannot provide a reliable reference for doctors. Unlike the improved LSTM model in Example 2, it can accurately predict the changes in postoperative feature quantities.

[0083] Although some data were integrated during the comprehensive evaluation, the correlation between the feature quantities was not considered, and the weighted sum was simply calculated. The weight setting lacked a basis. The doctor gave a "lower-middle" evaluation based on this rough result, which was too general and could not refine the rehabilitation recommendations based on the precise score as in Example 2. The subsequent treatment of the patient was poorly targeted.

[0084] 4. Results comparison For a period of time after surgery, the actual joint recovery was poor, and problems such as unstable condylar position and increased wear of the articular disc occurred, which were not effectively warned by the previous evaluation. Due to inaccurate rehabilitation guidance, the patient repeatedly adjusted the treatment plan, which was time-consuming and laborious, and increased pain, which had a great impact on life and work, in sharp contrast to the smooth recovery of the patient in Example 2 based on scientific evaluation.

[0085] Compared with Example 2, the patient under the traditional treatment mode of this comparative example had a tortuous treatment process and poor rehabilitation effect due to backward technical means and inaccurate evaluation, which once again confirmed the key role of accurate evaluation and scientific guidance of the present invention.

[0086] Comparative Example 3 1. Data collection and processing A 55-year-old male patient with a BMI of 29 and a history of bilateral temporomandibular joint disorder for 10 years was selected. His symptoms had recently worsened and he was scheduled for joint fusion surgery. An old 2.0T MRI scanner was used. Due to the performance limitations of the equipment, the image quality was poor, with a layer thickness of 3mm, a layer spacing of 1mm, a matrix of 220×220, and a field of view (FOV) of 15cm. The edges of structures such as the articular disc and condyle were blurred, and the details were difficult to distinguish. When collecting the medical history, only joint pain and limited mouth opening were focused on, and no in-depth understanding of the patient's long-term eating habits (such as a preference for hard food) and occupational characteristics (such as long-term desk work and poor neck posture that may affect the temporomandibular joint) and other factors were obtained.

[0087] Image denoising uses a simple low-pass filter, which seriously damages image details and blurs the joint structure. Without effective image enhancement, the original low-quality image has a more obvious disadvantage in subsequent processing, and the subtle features of the articular disc and condyle cannot be clearly distinguished, which is a world of difference from the high-quality image data foundation in Example 3.

[0088] 2. Feature extraction Manual segmentation of the articular disc and condyle, lack of operator experience, rough segmentation boundaries, and missing key contour information in many places. There is no standardized method for determining feature points, and only visual observation is used. For example, the positioning deviation of the condyle top can reach ±10 pixels, and the deviation of the midpoint of the posterior edge of the disc zone is ±8 pixels, resulting in a disc-condyle angle error of up to ±25°, which completely loses its reference value and is far from the accurate feature point positioning in Example 3.

[0089] The disc thickness measurement was random, with several points randomly selected for measurement without considering any directional correlation, resulting in a high degree of randomness. The MRI signal strength was determined based on subjective feelings only, without any quantitative analysis, and without any knowledge of the disc tissue characteristics and possible effusion, inflammation, etc., which was completely different from the comprehensive signal strength assessment in Example 3.

[0090] 3. Evaluation Method No model was used for postoperative prediction, and the process completely relied on the doctor's subjective assumptions. There was no forward-looking analysis of the patient's postoperative recovery process, such as vague statements such as "wait for the treatment to see the effect" and "it should get better slowly". There was a lack of scientific basis and quantitative indicator support, unlike Example 3, which used the improved LSTM model to predict postoperative joint status changes in advance.

[0091] The comprehensive evaluation is simply based on the patient's current symptoms and roughly collected information. The doctor gives a qualitative suggestion of "treat and observe first". There is no multi-layer MLP fusion of multi-source data to output an accurate score like in Example 3, which cannot provide direction for subsequent treatment planning, and the patient's recovery process is largely blind.

[0092] 4. Results comparison After a period of treatment, due to inaccurate early assessment, the patient's actual joint problems were not accurately treated. Although the joint pain was relieved, the problem of limited mouth opening was not improved, and even the joint snapping was aggravated, affecting the patient's eating and speech expression. Reexamination found that problems such as abnormal disc thickness were not taken seriously in the early stage, delaying the treatment opportunity, which was in sharp contrast to the patient's accurate recovery according to the scientific process in Example 3.

[0093] Compared with Example 3, the patient had a poor treatment effect under the backward treatment method of this comparative example due to the lack of accurate evaluation technology, which highlights the core advantages of the present invention in the entire process of diagnosis and treatment of the temporomandibular joint and provides reliable technical support for clinical practice.

[0094] The present invention compares three groups of comparative examples with the embodiments, and from the multi-dimensional perspective of the same patients and different patients, deeply analyzes the many drawbacks of traditional methods in the evaluation of temporomandibular joint repositioning surgery: rough data collection and processing, inaccurate feature extraction, backward evaluation model, and subjective and arbitrary comprehensive evaluation. However, the present invention, relying on rigorous processes and advanced technologies (such as CNN segmentation, improved LSTM model, multi-layer MLP fusion), accurately mines joint status information, outputs quantitative scores to guide rehabilitation, greatly improves the accuracy of postoperative evaluation and the effectiveness of treatment, and brings revolutionary changes to the rehabilitation of patients with temporomandibular joint disorders, and has broad clinical application prospects.

[0095] The above is only a preferred embodiment of the present invention. It should be pointed out that for ordinary technicians in this technical field, several improvements and modifications can be made without departing from the technical principles of the present invention. These improvements and modifications should also be regarded as the scope of protection of the present invention.

Claims

1. A method for evaluating the temporomandibular joint after reduction based on sagittal disc MRI images, characterized in that: The following steps are involved: S1: Data collection, including open-mouth MRI images, closed-mouth MRI images, and patient age, BMI, and gender information; S2: Image preprocessing, denoising and image enhancement of the acquired open-mouth and closed-mouth MRI images; S3: Image segmentation and feature point positioning: perform image segmentation of the articular disc and condyle on the preprocessed MRI image to determine the positions of the condylar apex, the condylar head center point, and the midpoint of the posterior edge of the posterior disc band; S4: Evaluation of disc protrusion angle and disc demarcation angle. The disc protrusion angle and disc demarcation angle are calculated based on the orbitoauricular plane coordinate system to determine the disc position and displacement. S5: Disc thickness assessment, measuring disc thickness in x / y direction and analyzing correlation to determine shape migration; S6: MRI signal intensity assessment, quantification of the gray value of the disc area to analyze the changes in effusion and tissue characteristics; S7: Condylar position assessment, condylar displacement is assessed by the joint space index; S8: Model prediction, using improved LSTM to predict the feature values ​​3, 6, and 12 months after surgery; S9: Correlation calculation, calculate the Pearson, Spearman, Kendall and distance correlation coefficients of the feature values; S10: Comprehensive evaluation, integrating MRI data, feature quantities, correlation coefficients and patient information, inputting into MLP to output health score α2.

2. The method for evaluating the temporomandibular joint after reduction based on sagittal disc MRI images according to claim 1, characterized in that: The image segmentation and feature point positioning are as follows: using a convolutional neural network to perform image segmentation of the articular disc and condyle on the preprocessed MRI image; The evaluation of the disc process angle and disc demarcation angle is as follows: the coordinate system is defined based on the orbitoauricular plane, the baseline for evaluating the disc process angle and disc demarcation angle is determined according to the positions of the determined condyle and articular disc related points, and then the disc process angle and disc demarcation angle are calculated, the position of the articular disc is determined by the disc process angle, and whether the articular disc is displaced anteriorly or laterally by the disc demarcation angle; The specific evaluation of articular disc thickness is as follows: on the segmented articular disc image, the dimensions of the articular disc in the x-direction and the y-direction are measured, and the correlation between the changes in the x-direction and the y-direction is calculated to determine the shape migration of the articular disc; MRI signal intensity assessment specifically includes: quantifying the grayscale value of the articular disc area in the MRI image to determine the signal intensity, analyzing the parameters of the signal intensity, setting the brightness threshold to determine the effusion in the articular disc area, and assisting in determining changes in the characteristics of the articular disc tissue based on changes in signal intensity; The condylar position assessment is as follows: measuring the posterior condylar space and the anterior joint space, using the anterior and posterior joint space base to calculate the joint space index, accurately indicating the displacement of the condyle relative to the central position, and assessing the condylar position status; The model prediction is as follows: the improved LSTM model is used to analyze and predict the five feature quantities extracted before and during surgery, namely, the disc angle, disc demarcation angle, disc thickness, disc MRI signal intensity, and condylar position, and the corresponding five feature quantities are obtained 3 months, 6 months, and 12 months after surgery to take into account the dynamic changes after surgery; The correlation calculation is specifically as follows: Pearson correlation coefficient, Spearman correlation coefficient, Kendall correlation coefficient and distance correlation analysis are calculated between the above five feature quantities to explore the internal connection between the feature quantities; The comprehensive evaluation is as follows: the original collected open and closed mouth MRI image data, the calculated feature quantities, the correlation coefficients between the feature quantities, and the individual information of the patient are integrated into a new data set, which is input into the multi-layer MLP. After being processed by the multi-layer MLP hidden layer, the compressed information variable is first output, and then further compressed to output the overall health score α2, where α2 belongs to [0.9, 0.8, 0.7, 0.6].

3. The method for evaluating the temporomandibular joint after reduction based on sagittal disc MRI images according to claim 2, characterized in that: The input of the forget gate of the improved LSTM model is extended to include the hidden states of the previous two steps and the current input, and the update formula of the forget gate is: , where W f and b f are the weight and bias parameters of the forget gate, is the sigmoid activation function, is the hidden state at the previous time step t-1, is the hidden state at the first two time steps t-2, The input feature vector at the current time step t.

4. The method for evaluating the temporomandibular joint after reduction based on sagittal disc MRI images according to claim 3, characterized in that: The improved LSTM model designs a new way to update the cell state, the formula is: , where W C and b C are the weights and bias parameters of candidate memory cells, is a candidate memory cell, which is a method for calculating candidate memory cells.

5. The method for evaluating the temporomandibular joint after reduction based on sagittal disc MRI images according to claim 4, characterized in that: The cell state update formula of the improved LSTM model is: , where C t - 1 is the cell state at the previous time step, is the output of the forget gate. This formula is the final cell state update rule, which is a calculation formula that combines the forget gate output with the candidate memory cell.

6. The method for evaluating the temporomandibular joint after reduction based on sagittal disc MRI images according to claim 5, characterized in that: The calculation of the output gate of the improved LSTM model remains unchanged. The output gate determines which information in the cell state will be output as the hidden state. The formula is: ,in and b o are the weight and bias parameters of the output gate, which are calculated for the output gate and determine which information in the cell state is output.

7. The method for evaluating the temporomandibular joint after reduction based on sagittal disc MRI images according to claim 6, characterized in that: The final hidden state of the improved LSTM model is determined by the output gate and the cell state, and the formula is: , this formula is for hidden state calculation, the hidden state is the output gate Acting on cell state The result is then passed to the next time step.

8. The method for evaluating the temporomandibular joint after reduction based on sagittal disc MRI images according to claim 2, characterized in that: The multi-layer MLP has 6 hidden layers, 50 nodes in the first three layers, 10 nodes in the fourth and fifth layers, 5 nodes in the sixth layer, and 1 node in the output.

9. The method for evaluating the temporomandibular joint after reduction based on sagittal disc MRI images according to claim 2, characterized in that: When the disc process angle is between 0° and -60°, the position of the articular disc is favorable for overcorrection; when the disc process angle is greater than -60°, the position of the articular disc is truly overcorrected; when the disc process angle is between 0° and 90°, the articular disc is in a stable position dependent on the occlusion plate; when the disc process angle exceeds 90°, the position of the articular disc is in a recurrent state.

10. The method for evaluating the temporomandibular joint after reduction based on sagittal disc MRI images according to claim 2, characterized in that: When the anterior-posterior angle between the tangent line of the posterior edge of the articular disc and the vertical line passing through the 12 o'clock direction of the condyle exceeds 10°, it is anterior displacement; the calculation formula of the joint space index is: , where P represents the posterior joint space and A represents the anterior joint space.

Images