Automatic determination of skeletal age from hand radiographs of children

1. Automatic determination of skeletal age from hand radiographs of children Image Science Institute Utrecht University C.A.Maas

2. Outline • Introduction • Automated procedures • preprocessing operations • segmentation of the hand • staging of the radius • Discussion • Conclusion

3. Introduction • Motivation • Development of the hand • Estimating the skeletal age • Greulich and Pyle • Tanner and Whitehouse

4. Project setup • Goal: • Invest possibilities for automating the skeletal age determination • Tasks: • preprocessing operations • segmentation of the hand • staging of the radius

5. Preprocessing operations • Rotation • Framing • upside-down check

6. Rotation • Radiograph • Gradient • Histogram -30° 60°

7. Framing and upside down • Pixel value left and right of vertical line • Horizontal projection for average intensity

8. Algorithm

9. Results • Rotation 99% • Framing • Vertical  92% • Horizontal  79% • upside down  100%

10. Segmentation of the hand • Statistical Shape Model of the hand • Manual segmentation • 49 fixed landmark points • 66 intermediate points • Represent shape by vector • x = (x1,y1,x2,y2,….x115,y115) • N=100

11. Model variations • Shapes is points in 230-D space • Principal Component Analysis • Mean and covariance are calculated 1 1 2 2

12. Model variations • 99% of shapes represented by 13 modes

13. Active Shape Model • Each landmark points has its local profile • Find best fit, smallest Mahalanobis distance • Adjust model based on new positions landmark points • Iterate at different resolutions

14. Demonstration of ASM

15. Active Shape Model • Starting position is essential for result • Best starting shape: • Generate starting shapes • Select on Mahalanobis distance

16. Results • Starting position: average distance Average shape 27.5 pixels Best starting position 11.0 pixels • Segmentation: good moderately- moderately- bad good bad 77% 15%4% 4%

17. Regions of Interest • Indicate ROIs on training images • Warp pointset to average shape • Calculate average positions of ROIs • Estimate positions of ROIs based on points in average shape

18. Staging of radius E F G H Rotate Translate Scale I

19. Extension 1: Region • Boxshaped • Compare boxes • Landmark points • Use landmark point of ASM • Circles with diameter of 40 pixels

20. Extension 2: Comparison • Average image • reference images • 12 reference images per stage

21. ? Classifiers • 17 features • Linear Discriminant Classifier • k- Nearest Neighbor classifier • Leave-one-out

22. Reclassification • Confusion matrix B C D E F G H I B 2 0 0 1 0 0 0 0 C 0 2 0 0 0 0 0 0 D 0 0 5 0 0 0 0 0 E 0 0 2 13 4 1 1 0 F 0 0 0 13 26 2 1 0 G 0 0 0 1 22 16 2 0 H 0 0 0 0 0 5 43 0 I 0 0 0 0 0 0 17 9 62% similar classified 97% within one stage difference

23. Results (1/2) • Semi-ASM versus ASM • Select 10 features from the 17 features • kNN classifier

24. Results (2/2) Region comparison correct within one classified stage error Box average 39% 89% Box reference 46% 95% 17 circles average 58% 98% 17 circles reference × × Second observer 62% 97%

25. Discussion • Preprocessing operations • robustness • Segmentation of the hand • self evaluation • Staging of the Radius • Good ASM for each ROI • Further steps • combine alle techniques • staging of all ROIs

26. Conclusion • Preprocessing operations perform good (99%) • Segmenting hand with ASM is successful (92%) • kNN classifier works good • 17 circles and reference images improve results • Computer close to human 62 %; 97 % versus 58 %; 98 % • Better training data, equal distribution

27. END