Cervical Spine Range of Motion Measurement Utilizing Image Analysis - VISAPP2022
0 likes310 views
This study developed a system to automatically measure cervical spine range of motion (CRoM) angles from cervical spine X-ray images using deep learning. The system used Mask R-CNN for image segmentation and measured angles between vertebrae similarly to manual methods. An evaluation found the average error was 3.5 degrees with a standard deviation of 2.8 degrees, comparable to measurements by residents. However, accuracy was poorer for the C1/C2 vertebrae. Future work will explore improving segmentation and developing computer-aided diagnosis of cervical issues.
![• A Study on Lung Region Detection Using Deep
Learning
• Mask R-CNN is more robust to the diversity.
4
[1] Uozumi, H., Matsubara, N., Teramoto, A., Niki, A., Honmoto, T., Kono, T., Saito, K., and Fujita, H. (2020). Lung region segmentation on pediatric chest x-rays using
mask r-cnn in japanese. Med Imag Tech, 38(3), 126–131.
2. Related Work: Image Segmentation
Output results from Mask R-CNN and U-net[1]](https://image.slidesharecdn.com/kanamatsuovisapp2022ver5-220208225227/85/Cervical-Spine-Range-of-Motion-Measurement-Utilizing-Image-Analysis-VISAPP2022-4-320.jpg)
![• This study used deep learning and image processing to
measure the thickness of the prevertebral soft tissue.
• This system proposed a method to diagnose swelling by
referring to the flow of swelling diagnosis by physicians.
5
[2] Y. H. Lee et al.: Learning Radiologist's Step-by-Step Skill for Cervical Spinal Injury Examination, IEEE, Vol.6, pp. 55492 – 55500(2018)
2. Related Work: Computer Aided Diagnosis
Distance measurement flow[2] Example of data set[2]](https://image.slidesharecdn.com/kanamatsuovisapp2022ver5-220208225227/85/Cervical-Spine-Range-of-Motion-Measurement-Utilizing-Image-Analysis-VISAPP2022-5-320.jpg)
![• One of the methods for object detection and
segmentation
• Instance segmentation is possible.
9
[3] K He, G Gkioxari, P Dollar et al., "Mask r-cnn", Proceedings of the IEEE international conference on computer vision, pp. 2961-2969,(2017).
5. Machine learning model: Mask R-CNN[3]
Mask R-CNN architecture](https://image.slidesharecdn.com/kanamatsuovisapp2022ver5-220208225227/85/Cervical-Spine-Range-of-Motion-Measurement-Utilizing-Image-Analysis-VISAPP2022-9-320.jpg)
Cervical Spine Range of Motion Measurement Utilizing Image Analysis - VISAPP2022
- 1. Cervical Spine Range of Motion Measurement Utilizing Image Analysis 17th International Joint Conference on Computer Vision, Imaging and Computer Graphics Theory and Applications. Kana Matsuo1) , Koji Fujita 2) , Takahumi Koyama 2) , Shingo Morishita2) , Yuta Sugiura1) 1) Keio University 2) Tokyo Medical and Dental University
- 2. CRoM: each cervical spine range of motion angle between C1 and C7. Measurement purpose: analyzing the mobility between vertebrae from cervical X-ray images. 2 1. Background: CRoM Definition C1 C2 C3 C4 C5 C6 C7
- 3. • One of the areas on which many imaging studies have been focused. • The CRoM measurement is burdensome for physicians. • Physicians in areas in which they do not specialize sometimes perform diagnostic. 3 1. Background: CRoM Measurement Example of CRoM measurement annotation
- 4. • A Study on Lung Region Detection Using Deep Learning • Mask R-CNN is more robust to the diversity. 4 [1] Uozumi, H., Matsubara, N., Teramoto, A., Niki, A., Honmoto, T., Kono, T., Saito, K., and Fujita, H. (2020). Lung region segmentation on pediatric chest x-rays using mask r-cnn in japanese. Med Imag Tech, 38(3), 126–131. 2. Related Work: Image Segmentation Output results from Mask R-CNN and U-net[1]
- 5. • This study used deep learning and image processing to measure the thickness of the prevertebral soft tissue. • This system proposed a method to diagnose swelling by referring to the flow of swelling diagnosis by physicians. 5 [2] Y. H. Lee et al.: Learning Radiologist's Step-by-Step Skill for Cervical Spinal Injury Examination, IEEE, Vol.6, pp. 55492 – 55500(2018) 2. Related Work: Computer Aided Diagnosis Distance measurement flow[2] Example of data set[2]
- 6. • Purpose • To implement a system that automates the measurement of CRoM angles from cervical spine X-ray images. • Our method • Our method used deep learning to estimate the cervical spine region from x-ray images. • Our system was implemented based on the manual measurement method. 6 3. Overview of System Our system’s display RoM (deg)
- 7. 7 4. System Design Flow of our system process
- 8. 8
- 9. • One of the methods for object detection and segmentation • Instance segmentation is possible. 9 [3] K He, G Gkioxari, P Dollar et al., "Mask r-cnn", Proceedings of the IEEE international conference on computer vision, pp. 2961-2969,(2017). 5. Machine learning model: Mask R-CNN[3] Mask R-CNN architecture
- 10. Learning Environment Learning data: 922 images Test data: 46 images Platform: Google colabratory Libraries: Keras, TensorFlow Hyperparameters Number of epochs: 100epoch Classification: Three-class Image size: 512x512px Batch size: 1 Learning coefficient: 0.001 Optimization method: Stochastic Gradient Descent (SGD) 10 5. Machine learning model: Parameters Labeling
- 11. • 97% of the total test data was able to be detected. • the IoU values of C1 and C2 were lower. 11 5. Machine learning model : Estimation Results IoU of each cervical spine Vertebra IoU C1 0.74 C2 0.83 C3 0.88 C4 0.88 C5 0.87 C6 0.86 C7 0.86 Average 0.85 ※IoU (Intersection over Union)= | 𝐴𝐴∩𝐵𝐵 𝐴𝐴∪𝐵𝐵 | 𝐴𝐴 ∩ 𝐵𝐵 𝐴𝐴 ∪ 𝐵𝐵
- 12. 12 6. CRoM Measurement: Edge Detection Extract the coordinates for creating a convex hull from the contour coordinates. 1 Approximate a convex hull. 2 The edge where the calculated midpoints are close to each other is extracted. 3 Extract of contour coordinates near the selected edge. 4 A line is drawn on these coordinates using the least- squares method to find the angles. 5 2 3 4 5
- 13. • Calculated from Tangent’s additive theorem by using Equation (1), where a and b are the slopes of the two lines. • tan 𝜃𝜃 = tan 𝛼𝛼−tan 𝛽𝛽 1+tan 𝛼𝛼 tan 𝛽𝛽 = 𝑎𝑎−𝑏𝑏 1+𝑎𝑎𝑎𝑎 (1) 13 6. CRoM Measurement: Angle Between 𝑦𝑦 = 𝑎𝑎𝑎𝑎 𝑦𝑦 = 𝑏𝑏𝑏𝑏 𝜃𝜃 tan 𝛼𝛼 = 𝑎𝑎 tan 𝛽𝛽 = 𝑏𝑏 angle between two straight lines
- 14. True value • 2 specialists: 23 patients (test data)×3 times (number of measurements) • The frequency of measurements: once a day Comparison data • 2 residents: 23 patients (test data)×3 times (number of measurements) • The frequency of measurements: once a day • Our system: 133/138 places (23 patients×6 places) • The remaining 5 places were not measured because the cervical regions were not segmented at all. 14 7. Overview of Experiments
- 15. • The accuracy of the proposed system was particularly poor for C1/2. • There was no difference in the overall mean error between the automatic and residents measurement. 15 Resident (deg) Automatic (deg) Vertebra Average error Standard deviation Average error Standard deviation C1/2 5.9 8.9 5.7 4.4 C2/3 2.9 2.3 4.0 3.6 C3/4 3.1 2.7 2.5 2.4 C4/5 3.3 2.7 2.6 2.1 C5/6 2.8 2.2 3.6 2.3 C6/7 2.6 1.9 2.5 2.2 Average 3.5 3.4 3.5 2.8 8. Experiment Results: Comparison Result Average error compared to the specialist measurement Cervical spine image
- 16. 16 8. Experiment Results: t-test Result • p ＞0.05 (t-test result): there is no a significant difference. Comparison of average error between resident measurements and automatic measurements
- 17. • In the automatic measurement, the standard deviation was 0. • Standard deviation scale: automatic＜specialist＜resident 17 8. Experiment Results: Standard Deviation Vertebra Standard deviation (deg) C12 5.4 C23 2.6 C34 2.9 C45 3.0 C56 2.7 C67 2.7 Average 3.2 Standard deviation of resident measurements Vertebra Standard deviation (deg) C12 4.4 C23 2.4 C34 2.9 C45 2.5 C56 2.7 C67 2.4 Average 2.9 Standard deviation of specialist measurements
- 18. • The accuracy of the proposed system was particularly poor for C1/2 between C1/2 and C6/7. • The IoU values of C1 and C2 were smaller than those of other cervical regions. • There was no significant difference in the overall mean. • In the automatic measurement, the standard deviation was 0. • That’s a common advantage of most computer vision methods: repeatability. 18 9. Discussion
- 19. Issue in the future • The accuracy of measurements was particularly poor for C1/2. Future work • Comparison other segmentation methods • U-net and K-means method • Develop a computer aided diagnosis system • Determine the cervical misalignment to estimate the defective cervical spine region 19 10. Issue and Future Work Image of the estimation of cervical vertebral alignment (Left: normal Right: abnormal)
- 20. 20 11. Summary Background A prompt and accurate diagnosis is required. Related work Medical Image Segmentation, Computer Aided Diagnosis Suggestion Use a image segmentation and reference to the diagnostic flow of a physician. Method Implement a system that automates the measurement of CRoM angles from cervical spine X-ray images. Evaluation Performed a two-sided t-test at the 5% level of significance. Results Average error: 3.5 deg Standard deviation: 2.8 deg Issues The accuracy of the proposed system was particularly poor for C1/2.