Method for detecting polyps in a three dimensional image volume

1. A method for detecting polyps in a three dimensional (3D) image volume of an anatomical structure comprising the steps of:

receiving a three dimensional (3D) image volume of an anatomical structure;

obtaining a set of candidate locations in the image volume;

for each candidate location, crop out three sub-volumes, wherein each sub-volume comprises a different scale and a plurality of voxels;

for each of the three sub-volumes, obtain two additional sub-volumes (for a total of nine sub-volumes) at different orientations by rotating each sub-volume in two different directions;

for each of the nine sub-volumes, apply a shape classifier thereto;

discard voxels of each of the nine sub-volumes that are not identified by the shape classifier as being on a surface of a polyp, and for each voxel of each of the nine sub-volumes that is identified by the shape classifier as being on the surface of a polyp, perform the following:

(a) compute a gradient direction for the voxel;

(b) if the gradient direction for the voxel falls into one of a plurality of orientations, apply a probability classifier cascade to the voxel;

(c) if the voxel passes the probability classifier cascade, its bounding box is remembered and its corresponding probability is outputted as a confidence measure therefor; and

(d) repeat (a-c) for each remaining voxel of each of the nine sub-volumes;

if bounding boxes exist after each of the nine sub-volumes has been analyzed, the candidate with the bounding box having the highest confidence measure is determined to be a polyp; and

outputting the image volume with at least one polyp identified therein.

2. The method of claim 1 wherein the probability classifier is an AdaBoost classifier.

3. The method of claim 1 wherein the probability classifier is a Probabilistic Boosting Tree (PBT) classifier.

4. The method of claim 1 wherein the step of applying a probability classifier to the voxel further comprises the steps of:

applying a plurality of filters to each of the nine sub-volumes.

5. The method of claim 1 wherein the plurality of filters are Haar filters.

6. The method of claim 5 wherein the filters are 3D Haar filters.

7. The method of claim 5 wherein the Haar filters correspond to multiple combinations of different sizes and aspect ratios of templates.

8. The method of claim 5 wherein the Haar filters are computed offline.

9. The method of claim 1 wherein the 3D image volume is a Computed Tomography (CT) image volume.

10. The method of claim 1 wherein the anatomical structure is a colon.

11. The method of claim 1 wherein the anatomical structure is a lung.

12. The method of claim 1 wherein the anatomical structure is a lymph node.

13. A system for determining if candidate locations identified in a three dimensional (3D) image volume of an anatomical structure correspond to a polyp, comprising:

a memory device for storing a program:

a processor in communication with the memory device, the processor operative with the program to:

receive a three dimensional (3D) image volume of an anatomical structure;

obtain a set of candidate locations in the image volume;

for each candidate location, crop out three sub-volumes, wherein each sub-volume comprises a different scale and a plurality of voxels;

for each of the three sub-volumes, obtain two additional sub-volumes (for a total of nine sub-volumes) at different orientations by rotating each sub-volume in two different directions;

for each of the nine sub-volumes, apply a shape classifier thereto;

discard voxels of each of the nine sub-volumes that are not identified by the shape classifier as being on a surface of a polyp, and for each voxel of each of the nine sub-volumes that is identified by the shape classifier as being on the surface of a polyp, the processor is further operative with the program to:

(a) compute a gradient direction for the voxel;

(b) if the gradient direction for the voxel falls into one of a plurality of orientations, apply a probability classifier cascade to the voxel;

(c) if the voxel passes the probability classifier cascade, its bounding box is remembered and its corresponding probability is outputted as a confidence measure therefor; and

(d) repeat (a-c) for each remaining voxel of each of the nine sub-volumes;

if bounding boxes exist after each of the nine sub-volumes has been analyzed, the candidate with the bounding box having the highest confidence measure is determined to be a polyp; and

output the image volume with at least one polyp identified therein.

14. The method of claim 13 wherein the probability classifier is a Probabilistic Boosting Tree (PBT) classifier.

15. The system of claim 13 wherein the processor is further operative with the program when applying a probability classifier to the voxel to apply a plurality of Haar filters to each of the nine sub-volumes.

16. The method of claim 15 wherein the Haar filters correspond to multiple combinations of different sizes and aspect ratios of templates.

17. The method of claim 15 wherein the Haar filters are computed offline.

18. The method of claim 13 wherein the 3D image volume is a Computed Tomography (CT) image volume.

19. The method of claim 13 wherein the anatomical structure is a colon.

20. The method of claim 13 wherein the anatomical structure is a lung.

21. The method of claim 13 wherein the anatomical structure is a lymph node.