Method and system for automatic detection and classification of coronary stenoses in cardiac CT volumes

1. A method for detecting coronary stenoses in a 3D CT volume, comprising:

extracting coronary vessel centerlines from the 3D CT volume;

estimating a lumen cross-section area based on the extracted coronary vessel centerlines using a trained regression function; and

detecting stenosis candidates based on the estimated lumen cross-section area,

wherein the step of estimating a lumen cross-section area based on the extracted coronary vessel centerlines using a trained regression function comprises:

estimating the lumen cross-section area at each of a plurality of points in the extracted coronary vessel centerlines using the trained regression function.

2. The method of claim 1 , further comprising, prior to the step of estimating a lumen cross-section area based on the extracted coronary vessel centerlines:

detecting non-vessel regions along the extracted coronary artery centerlines; and

removing the detected non-vessel regions from the extracted coronary vessel centerlines.

3. The method of claim 2 , wherein the step of detecting non-vessel regions along the extracted coronary artery centerlines comprises:

detecting the non-vessel regions using a trained non-vessel region detector based on rotation invariant features extracted at each of a plurality of points along the coronary vessel centerlines using a cylindrical sampling pattern.

4. The method of claim 3 , wherein the trained non-vessel region detector is trained using random forests based on rotation invariant features extracted at each of a plurality of annotated training sample points using a cylindrical sampling pattern.

5. The method of claim 1 , wherein the step of estimating the lumen cross-section area at each of a plurality of points in the extracted coronary vessel centerlines using the trained regression function comprises:

estimating the lumen cross-section area at each of the plurality of points based on rotation invariant features extracted the respective point using a cylindrical sampling pattern.

6. The method of claim 1 , wherein the step of detecting stenosis candidates based on the estimated lumen cross-section area comprises:

generating a base-line curve using one of a low-pass filter and a spline smoother;

generating a de-trended residual curve by subtracting the base-line curve from the estimated lumen cross-section area in a segment of the coronary vessel centerlines;

extracting local maxima and local minima of the de-trended residual curve; and

detecting a stenosis candidate in the segment of the coronary vessel based on the local maxima and local minima of the de-trended residual curve.

7. The method of claim 6 , wherein the step of detecting stenosis candidates based on the estimated lumen cross-section area further comprises:

smoothing the de-trended residual curve prior to the step of extracting local maxima and local minima of the de-trended residual curve.

8. The method of claim 6 , wherein the step of detecting stenosis candidates based on the estimated lumen cross-section area further comprises:

determining a grade for each detected stenosis candidate based on estimated lumen radius within the detected stenosis candidate.

9. The method of claim 1 , further comprising:

automatically classifying the detected stenosis candidates.

10. The method of claim 9 , wherein the step of classifying the detected stenosis candidates comprises:

comparing at least one feature of each detected stenosis candidate to a threshold to determine whether the stenosis candidate is significant;

discarding all stenosis candidates determined not to be significant; and

classifying each significant stenosis candidate as one of calcified, non-calcified, and mixed.

11. The method of claim 10 , wherein the step of comparing at least one feature of each detected stenosis candidate to a threshold to determine whether the stenosis candidate is significant comprises:

comparing a grade calculated for each stenosis candidate to a threshold;

if the grade for a stenosis candidate is greater than the threshold, the stenosis candidate is determined to be significant;

if the grade for a stenosis candidate is not greater than the threshold, the stenosis candidate is determined not to be significant.

12. The method of claim 10 , wherein the step of classifying each significant stenosis candidate as one of calcified, non-calcified, and mixed comprises:

determining a calcified probability score for each of a plurality of points in a significant stenosis candidate using a trained calcified stenosis detector;

determining a non-calcified probability score for each of the plurality of points in the significant stenosis candidate using a trained non-calcified stenosis detector; and

classifying the significant stenosis candidate as one of calcified, non-calcified, and mixed based on the calcified-probability scores and the non-calcified probability scores of the plurality of points in the significant stenosis candidate.

13. The method of claim 12 , wherein the step of classifying the significant stenosis candidate based on the calcified-probability scores and the non-calcified probability scores of the plurality of points in the significant stenosis candidate comprises:

classifying each of the plurality of points in the significant stenosis candidate as one of calcified and non-calcified based on the calcified probability score and the non-calcified probability score for the respective point; and

classifying the significant stenosis candidate as one of calcified, non-calcified, and mixed based on the relative number of the plurality of points classified as calcified and non-calcified in the significant stenosis candidate.

14. The method of claim 1 , wherein the 3D CT image volume is a computed tomography angiograph (CCTA) image volume.

15. An apparatus for detecting coronary stenoses in a 3D CT volume, comprising:

means for extracting coronary vessel centerlines from the 3D CT volume;

means for estimating a lumen cross-section area based on the extracted coronary vessel centerlines using a trained regression function; and

means for detecting stenosis candidates based on the estimated lumen cross-section area,

wherein the means for estimating a lumen cross-section area based on the extracted coronary vessel centerlines using a trained regression function comprises:

means for estimating the lumen cross-section area at each of a plurality of points in the extracted coronary vessel centerlines using the trained regression function.

16. The apparatus of claim 15 , further comprising:

means for detecting non-vessel regions along the extracted coronary artery centerlines; and

means for removing the detected non-vessel regions from the extracted coronary vessel centerlines.

17. The apparatus of claim 16 , wherein the means for detecting non-vessel regions along the extracted coronary artery centerlines comprises:

means for detecting the non-vessel regions using a trained non-vessel region detector based on rotation invariant features extracted at each of a plurality of points along the coronary vessel centerlines using a cylindrical sampling pattern.

18. The apparatus of claim 15 , wherein the means for detecting stenosis candidates based on the estimated lumen cross-section area comprises:

means for generating a base-line curve using one of a low-pass filter and a spline smoother;

means for generating a de-trended residual curve by subtracting the base-line curve from the estimated lumen cross-section area in a segment of the coronary vessel centerlines;

means for extracting local maxima and local minima of the de-trended residual curve; and

means for detecting a stenosis candidate in the segment of the coronary vessel based on the local maxima and local minima of the de-trended residual curve.

19. The apparatus of claim 15 , further comprising:

means for automatically classifying the detected stenosis candidates.

20. The apparatus of claim 19 , wherein the means for classifying the detected stenosis candidates comprises:

means for comparing at least one feature of each detected stenosis candidate to a threshold to determine whether the stenosis candidate is significant;

means for discarding all stenosis candidates determined not to be significant; and

means for classifying each significant stenosis candidate as one of calcified, non-calcified, and mixed.

21. The apparatus of claim 20 , wherein the means for classifying each significant stenosis candidate as one of calcified, non-calcified, and mixed comprises:

means for determining a calcified probability score for each of a plurality of points in a significant stenosis candidate using a trained calcified stenosis detector;

means for determining a non-calcified probability score for each of the plurality of points in the significant stenosis candidate using a trained non-calcified stenosis detector; and

means for classifying the significant stenosis candidate as one of calcified, non-calcified, and mixed based on the calcified-probability scores and the non-calcified probability scores of the plurality of points in the significant stenosis candidate.

22. A non-transitory computer readable medium encoded with computer executable instructions for detecting coronary stenoses in a 3D CT volume, the computer executable instructions defining steps comprising:

extracting coronary vessel centerlines from the 3D CT volume;

estimating a lumen cross-section area based on the extracted coronary vessel centerlines using a trained regression function; and

detecting stenosis candidates based on the estimated lumen cross-section area,

wherein the computer executable instructions defining the step of estimating a lumen cross-section area based on the extracted coronary vessel centerlines using a trained regression function comprise computer executable instructions defining the step of:

estimating the lumen cross-section area at each of a plurality of points in the extracted coronary vessel centerlines using the trained regression function.

23. The computer readable medium of claim 22 , further comprising computer executable instructions defining the steps of:

detecting non-vessel regions along the extracted coronary artery centerlines; and

removing the detected non-vessel regions from the extracted coronary vessel centerlines.

24. The computer readable medium of claim 23 , wherein the computer executable instructions defining the step of detecting non-vessel regions along the extracted coronary artery centerlines comprise computer executable instructions defining the step of:

detecting the non-vessel regions using a trained non-vessel region detector based on rotation invariant features extracted at each of a plurality of points along the coronary vessel centerlines using a cylindrical sampling pattern.

25. The computer readable medium of claim 22 , wherein the computer executable instructions defining the step of detecting stenosis candidates based on the estimated lumen cross-section area comprise computer executable instructions defining the steps of:

generating a base-line curve using one of a low-pass filter and a spline smoother;

generating a de-trended residual curve by subtracting the base-line curve from the estimated lumen cross-section area in a segment of the coronary vessel centerlines;

extracting local maxima and local minima of the de-trended residual curve; and

detecting a stenosis candidate in the segment of the coronary vessel based on the local maxima and local minima of the de-trended residual curve.

26. The computer readable medium of claim 22 , further comprising computer executable instructions defining the step of:

automatically classifying the detected stenosis candidates.

27. The computer readable medium of claim 26 , wherein the computer executable instructions defining the step of classifying the detected stenosis candidates comprise computer executable instructions defining the steps of:

comparing at least one feature of each detected stenosis candidate to a threshold to determine whether the stenosis candidate is significant;

discarding all stenosis candidates determined not to be significant; and

classifying each significant stenosis candidate as one of calcified, non-calcified, and mixed.

28. The computer readable medium of claim 27 , wherein the computer executable instructions defining the step of classifying each significant stenosis candidate as one of calcified, non-calcified, and mixed comprise computer executable instructions defining the steps of:

determining a calcified probability score for each of a plurality of points in a significant stenosis candidate using a trained calcified stenosis detector;

determining a non-calcified probability score for each of the plurality of points in the significant stenosis candidate using a trained non-calcified stenosis detector; and

classifying the significant stenosis candidate as one of calcified, non-calcified, and mixed based on the calcified-probability scores and the non-calcified probability scores of the plurality of points in the significant stenosis candidate.