System and method for automated detection of lung nodules in medical images

1. A method for automatic detection of lung nodules, comprising:

inputting data from a computed tomography (CT) image;

segmenting the image to acquire one or more lung regions;

intensity thresholding the one or more lung regions to generate a mask region comprising one or more high-intensity regions corresponding to anatomical structures within the one or more lung regions;

computing a Euclidean distance map of the mask region;

performing watershed segmentation of the Euclidean distance map to generate one or more sub-regions;

identifying a seed point for each sub region of the one or more sub-regions;

growing one or more candidate regions from the seed point of each sub-region; and

classifying one or more candidate regions as a lung nodule based on one or more geometric features of the one or more candidate regions.

2. A method as recited in claim 1 , wherein segmenting the image comprises:

segmenting a chest wall region of the image by intensity thresholding bright voxels in the CT image and selecting the largest connected 3D component; and

segmenting a lung region from the chest wall region by intensity thresholding dark voxels in the CT image and selecting the largest connected 3D component.

3. A method as recited in claim 2 , wherein segmenting the image further comprises computing the convex hull of the segmented lung region.

4. A method as recited in claim 1 , wherein intensity thresholding the one or more lung regions is performed at a first threshold value for detection of solid nodule candidate regions.

5. A method as recited in claim 1 , wherein intensity thresholding the one or more lung regions is performed at a second threshold value for detection of ground glass nodule candidate regions.

6. A method as recited in claim 1 , wherein Gaussian smoothing is applied to the Euclidean distance map prior to watershed segmentation.

7. A method as recited in claim 1 , wherein the one or more sub-regions generated from watershed segmentation are contiguous with local maxima in the Euclidean distance map.

8. A method as recited in claim 7 , wherein identifying a seed point for each sub-region comprises identifying a voxel having a maximum Euclidean distance map value within each sub-region as the seed corresponding to the sub-region.

9. A method as recited in claim 8 , wherein growing one or more candidate regions comprises:

generating a cubic region of interest centered around each seed point; and

including continuous voxels within the region of interest as voxels in a candidate region;

the included voxels having Euclidean distance map values falling within a predetermined threshold range.

10. A method as recited in claim 9 , further comprising dilating each of the one or more grown candidate regions.

11. A method as recited in claim 1 , wherein classifying one or more candidate regions comprises:

calculating a volume of the candidate region; and

identifying the candidate region as a nodule as a function of the calculated volume falling within a threshold range.

12. A method as recited in claim 11 , wherein classifying one or more candidate regions further comprises:

calculating a sphericity of the candidate region;

wherein the sphericity is calculated as a ratio of the calculated volume of the candidate region to a volume of a minimum sphere centered about a centroid of the candidate volume, and

identifying the candidate region as a nodule as a function of the calculated sphericity exceeding a threshold value.

13. A method as recited in claim 1 , further comprising calculating a volume of an identified lung nodule by summing volumes of each voxel in the identified lung nodule.

14. A method as recited in claim 1 , further comprising calculating a diameter of an identified lung nodule.

15. A method as recited in claim 14 , wherein the diameter is calculated by:

dividing the lung nodule into a plurality of 2D regions of interest in one or more planes;

extracting boundary points for each 2D region of interest;

identifying a maximum diameter from the boundary points in each 2D region of interest; and

identifying a largest maximum diameter from each of the planes as the nodule diameter.

16. A method as recited in claim 1 , further comprising overlaying an image of the identified lung nodule over the CT image.

17. A system for automatically segmenting a computed tomography (CT) image of a patient's lung, comprising:

a processor;

programming executable on the processor and configured for:

inputting data from the CT image;

segmenting the image to acquire one or more lung regions;

intensity thresholding the one or more lung regions to generate a mask region comprising one or more high-intensity regions corresponding to anatomical structures within the one or more lung regions;

computing a Euclidean distance map of the mask region;

performing watershed segmentation of the Euclidean distance map to generate one or more sub-regions;

identifying a seed point for each sub region of the one or more sub-regions;

growing one or more candidate regions from the seed point of each sub-region; and

classifying one or more candidate regions as a lung nodule based on one or more geometric features of the one or more candidate regions.

18. A system as recited in claim 17 , wherein segmenting the image comprises:

segmenting a chest wall region of the image by intensity thresholding bright voxels in the CT image and selecting the largest connected 3D component; and

segmenting a lung region from the chest wall region by intensity thresholding dark voxels in the CT image and selecting the largest connected 3D component.

19. A system as recited in claim 18 , wherein segmenting the image further comprises computing the convex hull of the segmented lung region.

20. A system as recited in claim 17 , wherein intensity thresholding the one or more lung regions is performed at a first threshold value for detection of solid nodule candidate regions.

21. A system as recited in claim 17 , wherein intensity thresholding the one or more lung regions is performed at a second threshold value for detection of ground glass nodule candidate regions.

22. A system as recited in claim 17 , wherein Gaussian smoothing is applied to the Euclidean distance map prior to watershed segmentation.

23. A system as recited in claim 17 , wherein the one or more sub-regions generated from watershed segmentation are contiguous with local maxima in the Euclidean distance map.

24. A system as recited in claim 23 , wherein identifying a seed point for each sub region comprises indentifying a voxel having a maximum Euclidean distance map value of each sub-region as the seed corresponding to the sub-region.

25. A system as recited in claim 24 , wherein growing one or more candidate regions comprises:

generating a cubic region of interest centered around each seed point; and

including continuous voxels within the region of interest as voxels in a candidate region;

the included voxels having Euclidean distance map values falling within a predetermined threshold range.

26. A system as recited in claim 25 , further comprising:

dilating each of the one or more grown candidate regions.

27. A system as recited in claim 17 , wherein classifying one or more candidate regions comprises:

calculating a volume of the candidate region; and

identifying a candidate region as a nodule as a function of the calculated volume falling within a threshold range.

28. A system as recited in claim 27 , wherein classifying one or more candidate regions further comprises:

calculating a sphericity of a candidate region;

wherein the sphericity is calculated as a ratio of the calculated volume of a candidate region to a volume of a minimum sphere centered about a centroid of the candidate volume, and

identifying a candidate region as a nodule as a function of the calculated sphericity exceeding a threshold value.

29. A system as recited in claim 27 , further comprising:

calculating a volume of an identified lung nodule by summing volumes of each voxel in the identified lung nodule.

30. A system as recited in claim 17 , further comprising calculating a diameter of an identified lung nodule.

31. A system as recited in claim 30 , wherein the diameter is calculated by:

dividing the lung nodule into a plurality of 2D regions of interest in one or more planes;

extracting boundary points for each 2D region of interest;

identifying a maximum diameter from the boundary points in each 2D region of interest; and

identifying a largest maximum diameter from each of the planes as the nodule diameter.

32. A system as recited in claim 17 , further comprising overlaying an image of the identified lung nodule over the CT image.