Knee osteoarthritis assessment method based on machine learning
By using machine learning technology to extract and segment meniscus signal features from MRI images, the complexity of meniscus diagnosis in MRI images was solved, accurate diagnosis of knee osteoarthritis was achieved, and diagnostic efficiency and accuracy were improved.
Patent Information
- Application Number
- CN202510608759.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-05-13
- Publication Date
- 2025-09-26
AI Technical Summary
Existing technologies cannot effectively utilize the meniscus signal features in MRI images for accurate diagnosis of knee osteoarthritis, and traditional image segmentation methods are difficult to achieve ideal meniscus segmentation effects.
A machine learning method was used to accurately extract meniscus signal features, segment the knee cartilage and meniscus using T1- and T2-weighted images, train an automatic meniscus segmentation model, perform three-category classification of meniscus images, calculate the meniscus spatial specific signal index, and perform diagnostic evaluation in combination with the K nearest neighbor model.
It achieves automatic and precise segmentation of the meniscus, improves the diagnostic accuracy of knee osteoarthritis, provides a more objective and accurate diagnostic basis, and assists orthopedic surgeons in making more accurate clinical judgments.
Smart Images
Figure CN120707465A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of machine learning technology, and in particular to a method for evaluating knee osteoarthritis based on machine learning. Background Art
[0002] Knee osteoarthritis (KOA) is a common joint disease and is currently the eleventh leading cause of disability worldwide. The Kenneth-Lawrence grading system, based on knee X-ray images, is the gold standard for clinical diagnosis of KOA. However, X-rays only provide clear images of the skeletal system and cannot provide relevant information about connective tissues such as the meniscus and cartilage. Furthermore, the Kenneth-Lawrence grading system relies on the subjective experience of clinicians, and different clinicians may have inconsistent judgments.
[0003] As a non-invasive diagnostic technique, magnetic resonance imaging (MRI) can provide high-resolution images of the interior of the joint, clearly showing soft tissue structures such as the meniscus and articular cartilage within the knee joint. This information is crucial for the diagnosis and evaluation of KOA. With the advancement of medical image processing technology, machine learning is increasingly being used in the diagnosis and treatment evaluation of KOA. For example, machine learning can detect and classify knee joint tissue injuries and extract deep-level features, thereby improving diagnostic accuracy and efficiency.
[0004] Previous studies have shown that meniscus injury is an important indicator for KOA diagnosis, and the extraction of its signal features plays a key role in injury assessment. However, the diversity and irregularity of meniscus injuries complicate signal feature extraction. Furthermore, the signals presented by the meniscus in MRI images are complex and variable, making it difficult to achieve ideal segmentation using traditional image segmentation methods.
[0005] Therefore, providing a knee osteoarthritis assessment method based on machine learning has important clinical significance for the accurate diagnosis of knee osteoarthritis. Summary of the Invention
[0006] To address the above technical issues, the present invention proposes a method for assessing knee osteoarthritis based on machine learning. By accurately extracting and analyzing meniscus signal characteristics, it provides a more objective and accurate diagnosis basis for knee osteoarthritis.
[0007] The technical problem to be solved by the present invention is achieved by adopting the following technical solutions:
[0008] A method for assessing knee osteoarthritis based on machine learning, comprising the following steps:
[0009] Step 1: Acquire T1- and T2-weighted images of the knee joints of patients with knee osteoarthritis;
[0010] Step 2: Manually segment the knee cartilage in T1-weighted images and measure the knee cartilage volume;
[0011] Step 3: crop the meniscus image from the T2-weighted image and manually segment it to obtain labeled data to construct an image dataset;
[0012] Step 4: training a meniscus automatic segmentation model to segment the meniscus in the meniscus image;
[0013] Step 5: performing three-class classification on the meniscus pixels in the middle five layers of the meniscus image and calculating the meniscus spatial specific signal index;
[0014] Step 6. Systematically evaluate the performance of meniscus spatial-specific signal index in the diagnosis of knee osteoarthritis.
[0015] Preferably, in step 1, image quality and data enhancement processing is performed on T1 and T2 weighted images of the knee joint of patients with knee osteoarthritis, wherein the T1 weighted image is used to measure the volume of knee cartilage and the T2 weighted image is used to extract meniscus signal characteristics.
[0016] Preferably, in step 2, the knee cartilage boundary is manually segmented in the T1-weighted image, and the knee cartilage volume is measured, and the knee cartilage volume is one of the commonly used KOA diagnostic indicators.
[0017] Preferably, in step 3, the Hough circle detection method is used to locate the femur, the circle whose center is located in the lower half of the image is removed, the Euclidean distance Eu.dst between the center of the remaining circle and the image center is calculated, and the circle with the smallest Euclidean distance is selected as the target circle. The center of the target circle is used as the reference point to crop the meniscus image from the T2-weighted image, select the middle five slices of the meniscus image, manually segment the meniscus to obtain meniscus annotation data and create an image dataset; the Euclidean distance calculation formula is as follows:
[0018]
[0019] Among them, (x1, y1) represents the coordinates of the center of the remaining circle in the image, and (x2, y2) represents the coordinates of the center of the image.
[0020] Preferably, in step 4, the image dataset is used to train the meniscus automatic segmentation model, and the meniscus automatic segmentation model with the best performance is selected to segment the lateral and medial menisci in the meniscus image.
[0021] Preferably, in step 5, the grayscale value of the meniscus image is normalized to scale its range to between 0 and 1, and the formula is as follows:
[0022]
[0023] Among them, y represents the normalized grayscale value, x represents the original grayscale value, min represents the minimum grayscale value of all pixels in the image, and max represents the maximum grayscale value of all pixels in the image.
[0024] Preferably, a clustering algorithm is used to classify the pixels of the meniscus into three categories, which are recorded as the first category C1, the second category C2, and the third category C3. The first category C1 has the lowest signal intensity, the third category C3 has the highest signal intensity, and the second category C2 is between the first category C1 and the third category C3. The classified meniscus images are visualized.
[0025] Preferably, the volumes occupied by pixels of three signal intensities in the five slices and the total volume of the meniscus are calculated respectively, and the calculation formula is as follows:
[0026] Volume=Pixel Area ×Slice Thickness ×N Pixels (III)
[0027] Among them, Pixel Area is the area occupied by each pixel on the image plane, Slice Thickness is the distance between adjacent slices, N Pixels is the total number of pixels in this class.
[0028] Preferably, based on the calculated volume, the ratio of the volume occupied by pixels of the three signal intensities of the lateral and medial meniscus in the five slices to the total volume of the meniscus is calculated respectively to obtain six meniscus spatial specific signal indices, namely L-MSS-C1, L-MSS-C2, L-MSS-C3, M-MSS-C1, M-MSS-C2, and M-MSS-C3.
[0029] Preferably, in step six, the data set is randomly divided into a training set and a test set, and the grid search cross-validation technique and the five-fold cross-validation technique are used to calculate the five indicators of F1 score, recall rate, precision, accuracy and area under the curve. The knee cartilage volume, six MSS indices and their combinations are trained using the K nearest neighbor model to generate the above five indicators. Among them, the combination is as follows: L-MSS represents the combination of three MSS indices in the lateral meniscus, M-MSS represents the combination of three MSS indices in the medial meniscus, and LM-MSS represents the combination of all six MSS indices in the lateral and medial menisci. The five indicators obtained by comparing the six meniscus spatial specific signal indices, L-MSS, M-MSS, LM-MSS, and knee cartilage volume are evaluated to evaluate the best MSS index or MSS index combination and the best k value for diagnosing knee osteoarthritis.
[0030] The beneficial effects of the present invention are:
[0031] This paper provides a machine learning-based method for assessing knee osteoarthritis. It first uses machine learning to automatically and accurately segment the meniscus. It then classifies the meniscus pixels into three categories, calculating the ratio of the volume of each of the three signal intensity pixels to the total volume of the meniscus to obtain the MSS index. The K-nearest neighbor model is then used to systematically evaluate the performance of the meniscus spatially specific signal index in diagnosing KOA. This method, which assesses the extent of meniscus damage in the knee joint to diagnose knee osteoarthritis, can effectively assist orthopedic surgeons in making more accurate clinical decisions. BRIEF DESCRIPTION OF THE DRAWINGS
[0032] The present invention will be further described below with reference to the accompanying drawings and embodiments:
[0033] Figure 1 Flowchart of the present invention;
[0034] Figure 2 An example of meniscus segmentation and signal classification for a KOA patient, according to an embodiment of the present invention. (a) shows the femur localization (cyan circle) and a cropped meniscus image (pink rectangle) in a T2-weighted MRI. (b) shows a schematic diagram of meniscus segmentation and signal classification. (c) and (d) show magnified images of the lateral and medial menisci. (e) and (f) show reconstructed images of the lateral and medial menisci. DETAILED DESCRIPTION
[0035] In order to make the technical means, creative features, objectives and effects achieved by the present invention easier to understand, the present invention is further described below with reference to the accompanying drawings and embodiments.
[0036] like Figure 1 As shown, a method for evaluating knee osteoarthritis based on machine learning includes the following steps:
[0037] Step 1: Acquire T1- and T2-weighted images of the knee joints of patients with knee osteoarthritis.
[0038] In this example, MRI images of the knee joints of 90 patients with knee osteoarthritis were collected, and the images were weighted averaged using the Gaussian filtering method to enhance image quality. The geometric transformation method was used to increase data intensity to avoid changes in tissue signal intensity due to color transformation, thereby losing important diagnostic information. It should be noted that the scanning equipment involved in this example is a 3.0T GE Signa HDxt scanner (General Electric Healthcare, Little Chalfont, UK) and a commercial transmission / reception coil. The scanning sequence and parameters are:
[0039] (1) T1-weighted fat-suppressed three-dimensional spoiled gradient echo sequence, flip angle 30°, repetition time 31 ms, echo time 6.71 ms, field of view 16 × 16 cm², matrix 512 × 512, slice thickness 1.5 mm;
[0040] (2) T2-weighted fast spin echo fat-saturated two-dimensional sequence, flip angle 90°, repetition time 3067 milliseconds, echo time 112 milliseconds, field of view 16×16 square centimeters, matrix 512×512, slice thickness 2 mm, and slice gap 0.18 mm.
[0041] Through the above embodiment, the image quality and data enhancement processing of T1 and T2 weighted images are performed by Gaussian filtering method and geometric transformation method, wherein T1 weighted image is used to measure knee cartilage volume, and T2 weighted image is used to extract meniscus signal characteristics.
[0042] Step 2: Manually segment the knee cartilage in T1-weighted images and measure the knee cartilage volume.
[0043] In T1-weighted magnetic resonance imaging (MRI) images, knee articular cartilage was manually segmented using Osiris software, and the knee articular cartilage volume was measured.
[0044] Step 3: Crop the meniscus image from the T2-weighted image and manually segment it to obtain labeled data to construct an image dataset.
[0045] The meniscus is located between the femur and tibia. The edge of the femur is arc-shaped. Hough circle detection is used to accurately locate the femur. The femur is typically located in the upper half of the image, so circles centered in the lower half are excluded. The Euclidean distance (Eu.dst) between the center of each remaining circle and the image center is calculated. The circle with the smallest Euclidean distance is selected as the target circle. The Euclidean distance calculation formula is as follows:
[0046]
[0047] Among them, (x1, y1) represents the coordinates of the center of the remaining circle in the image, and (x2, y2) represents the coordinates of the center of the image.
[0048] Taking the center of the target circle as the reference point, extend 40% of the width of the original T2-weighted image to the left and right, and extend 20% of the height of the original T2-weighted image vertically downward to form a rectangular frame covering the main area of the meniscus. The meniscus image is cut from the T2-weighted image, as shown in the figure. Figure 2 As shown, Figure 2 (a) T2-weighted image showing the femur (cyan circle) and cropped meniscus image (pink rectangle). The middle five layers of the meniscus image were selected and manually segmented to obtain annotated meniscus data and create an image dataset.
[0049] Step 4: Train the meniscus automatic segmentation model to segment the meniscus in the meniscus image.
[0050] The image dataset was randomly split into training, validation, and test sets in an 8:1:1 ratio. Five deep convolutional neural network (CNN) models, including U-Net, U-Net++, DenseUnet, Swin-Unet, and TransUnet, were trained in the PyCharm Community Edition 2021 development environment. The segmentation models were comprehensively evaluated using metrics such as the Dice coefficient, intersection over union (IoU), sensitivity, 95% Hausdorff distance (HD95), and mean symmetric surface distance (ASSD). Table 1 shows the performance of these five deep CNN models in segmenting the meniscus, with the TransUnet model achieving the best segmentation performance.
[0051] Table 1 Performance of five deep CNN models in meniscus segmentation
[0052]
[0053] A patient with knee osteoarthritis was selected as an example. The TransUnet model meniscus automatic segmentation model was used to segment the lateral and medial menisci in the meniscus image. The segmentation results are shown in the figure below. Figure 2 (b).
[0054] Step 5: Classify the meniscus pixels in the middle five layers of the meniscus image into three categories and calculate the meniscus spatial specific signal index.
[0055] Normalize the grayscale value of the meniscus image and scale it to between 0 and 1 to ensure data consistency. The formula is as follows:
[0056]
[0057] Among them, y represents the normalized grayscale value, x represents the original grayscale value, min represents the minimum grayscale value of all pixels in the image, and max represents the maximum grayscale value of all pixels in the image.
[0058] Then, the K-means clustering algorithm is used to classify the pixels of the meniscus into three categories, which are recorded as the first category C1, the second category C2, and the third category C3. The first category C1 has the lowest signal intensity, the third category C3 has the highest signal intensity, and the second category C2 is between the first category C1 and the third category C3. The classified meniscus images are visualized. The classification results are shown as follows: Figure 2 (bd) shows that blue represents C1, green represents C2, and red represents C3. The classified meniscus pixels are reconstructed in three-dimensional space using the Visualization Toolkit, as shown in Figure 2 (ef) shown.
[0059] The volume occupied by pixels of the three signal intensities in the five slices and the total volume of the meniscus were calculated respectively. The calculation formula is as follows:
[0060] Volume=Pixel Area ×Slice Tickness ×N Pixels (III)
[0061] Among them, Pixel Area is the area occupied by each pixel on the image plane, Sice Thickness is the distance between adjacent slices, N Pixels is the total number of pixels in this class.
[0062] Based on these calculated volumes, the ratios of the volumes of pixels with three signal intensities in the lateral and medial menisci to the total volume of the meniscus were calculated for each of the five slices, yielding six meniscus spatially specific signal indices. For a patient with knee osteoarthritis, the meniscus spatially specific signal indices were: M-MSS-C1 = 0.89, M-MSS-C2 = 0.87, M-MSS-C3 = 0.02, L-MSS-C1 = 0.93, L-MSS-C2 = 0.05, and L-MSS-C3 = 0.01.
[0063] Step 6. Systematically evaluate the performance of the meniscus space-specific signal index in diagnosing knee osteoarthritis.
[0064] Using KLG-based KOA diagnosis as the gold standard and the K-nearest neighbor model, a common prediction method in machine learning, the performance of the MSS index in diagnosing KOA was systematically evaluated. Ninety subjects were randomly divided into training and test sets in a 7:3 ratio. Grid search cross-validation and five-fold cross-validation were used to calculate the F1 score, recall, precision, accuracy, and area under the curve.
[0065] A k-nearest neighbor model was used to train the five indices described above using knee cartilage volume, MSS indices, and their three combinations, obtained from segmentation of knee T1-weighted images. The specific combinations included L-MSS, representing the combination of the three MSS indices in the lateral meniscus; M-MSS, representing the combination of the three MSS indices in the medial meniscus; and LM-MSS, representing the combination of all six MSS indices in the lateral and medial menisci. Comparisons of the five indices derived from the six MSS indices, the three combinations, and knee cartilage volume revealed that M-MSS-C1 achieved an F1 score and recall of 0.85 and 0.87, respectively. Precision, accuracy, and area under the curve (AUC) for M-MSS were 0.84, 0.78, and 0.84, respectively. Precision, accuracy, and AUC for M-MSS-C1 were 0.83, 0.75, and 0.81, respectively. Table 2 shows the performance of the six MSS indices and their three combinations for diagnosing knee osteoarthritis using the k-nearest neighbor model and lists the optimal k value for each index or combination.
[0066] Table 2 Performance of six MSS indices and their three combinations in diagnosing knee osteoarthritis with cartilage volume
[0067]
[0068] Based on the above analysis, M-MSS has the best diagnostic effect for knee osteoarthritis among these indices.
[0069] The basic principles, main features, and advantages of the present invention are shown and described above. Those skilled in the art should understand that the present invention is not limited to the foregoing embodiments. The foregoing embodiments and description merely illustrate the principles of the present invention. Various changes and modifications may be made to the present invention without departing from the spirit and scope of the present invention. Such changes and modifications are intended to fall within the scope of the present invention. The scope of protection claimed in the present invention is defined by the appended claims and their equivalents.
Claims
1. A method for assessing knee osteoarthritis based on machine learning, characterized by: The following steps are involved: Step 1: Acquire T1- and T2-weighted images of the knee joints of patients with knee osteoarthritis; Step 2: Manually segment the knee cartilage in T1-weighted images and measure the knee cartilage volume; Step 3: crop the meniscus image from the T2-weighted image and manually segment it to obtain labeled data to construct an image dataset; Step 4: training a meniscus automatic segmentation model to segment the meniscus in the meniscus image; Step 5: performing three-class classification on the meniscus pixels in the middle five layers of the meniscus image and calculating the meniscus spatial specific signal index; Step 6. Systematically evaluate the performance of meniscus spatial-specific signal index in the diagnosis of knee osteoarthritis.
2. The method for assessing knee osteoarthritis based on machine learning according to claim 1, characterized in that: In step 1, the image quality and data enhancement processing is performed on the T1 and T2 weighted images of the knee joints of patients with knee osteoarthritis.
3. The method for assessing knee osteoarthritis based on machine learning according to claim 1, characterized in that: In step 2, the knee cartilage boundary was manually segmented in the T1-weighted images and the knee cartilage volume was measured.
4. The method for assessing knee osteoarthritis based on machine learning according to claim 1, wherein: In step 3, the Hough circle detection method is used to locate the femur. The circle with the center in the lower half of the image is removed. The Euclidean distance Eu.dst between the center of the remaining circle and the image center is calculated. The circle with the smallest Euclidean distance is selected as the target circle. The center of the target circle is used as the reference point to crop the meniscus image from the T2-weighted image. The middle five slices of the meniscus image are selected and the meniscus is manually segmented to obtain meniscus annotation data and create an image dataset. The Euclidean distance calculation formula is as follows: Among them, (x1, y1) represents the coordinates of the center of the remaining circle in the image, and (x2, y2) represents the coordinates of the center of the image.
5. The method for assessing knee osteoarthritis based on machine learning according to claim 4, characterized in that: In step 4, the image dataset is used to train the meniscus automatic segmentation model, and the meniscus automatic segmentation model with the best performance is selected to segment the lateral and medial menisci in the meniscus image.
6. The method for assessing knee osteoarthritis based on machine learning according to claim 5, characterized in that: In step 5, the grayscale value of the meniscus image is normalized to scale its range to between 0 and 1. The formula is as follows: Among them, y represents the normalized grayscale value, x represents the original grayscale value, min represents the minimum grayscale value of all pixels in the image, and max represents the maximum grayscale value of all pixels in the image.
7. The method for assessing knee osteoarthritis based on machine learning according to claim 6, characterized in that: In step five, a clustering algorithm is used to classify the pixels of the meniscus into three categories, namely the first category C1, the second category C2, and the third category C3. The first category C1 has the lowest signal intensity, the third category C3 has the highest signal intensity, and the second category C2 is between the first category C1 and the third category C3. The classified meniscus images are visualized.
8. The method for assessing knee osteoarthritis based on machine learning according to claim 7, characterized in that: The volumes occupied by pixels of the three signal intensities in the five slices and the total volume of the meniscus were calculated respectively. The calculation formulas are as follows: Volume=Pixel Area ×Slice Thickness ×N Pixels (III) Among them, Pixel Area is the area occupied by each pixel on the image plane, Slice Th ickness is the distance between adjacent slices, N Pixels is the total number of pixels in this class.
9. The method for assessing knee osteoarthritis based on machine learning according to claim 8, characterized in that: Based on the calculated volumes, the ratios of the volumes occupied by pixels of the three signal intensities of the lateral and medial menisci to the total volume of the meniscus in the five slices were calculated respectively, and six meniscus spatial-specific signal indices were obtained, namely L-MSS-C1, L-MSS-C2, L-MSS-C3, M-MSS-C1, M-MSS-C2, and M-MSS-C3.
10. The method for assessing knee osteoarthritis based on machine learning according to claim 9, characterized in that: In step 6, the dataset was randomly divided into a training set and a test set with a ratio of 7:
3. Grid search cross-validation and five-fold cross-validation techniques were used to calculate the five indicators of F1 score, recall rate, precision, accuracy, and area under the curve. The K-nearest neighbor model was used to train the knee cartilage volume, six MSS indices, and their combinations to generate the above five indicators. Among them, the combinations are as follows: L-MSS represents the combination of the three MSS indices in the lateral meniscus, M-MSS represents the combination of the three MSS indices in the medial meniscus, and LM-MSS represents the combination of all six MSS indices in the lateral and medial menisci. The five indices obtained by comparing the six meniscus spatial-specific signal indices, L-MSS, M-MSS, LM-MSS, and knee articular cartilage volume were evaluated to evaluate the optimal MSS index or MSS index combination and the optimal k value for diagnosing knee osteoarthritis.