Segmentation of volumetric images for accurate distinction of biologically significant entities

1. Segmentation of volumetric images for accurate distinction of biologically significant entities Ryan Green

2. Contents • Problems with volumetric images • Current approaches to image analysis • Flood-Fill • Parallelised RAM-efficient Flood-Fill • Conclusion

3. Problems with volumetric images • Difficult to determine biologically significant entities in 2D • Difficult to visualise in 3D • Large processor and RAM requirements

4. Magnetic Resonance Images Standard MR Images in 2 and 3 Dimensions

5. How images are currently analysed • Manual 2D visual analysis • Manual 3D visual analysis via transparencies according to tissue intensity • Automated/Semi-Automated Image segmentation via Fuzzy C-Means (FCM) and extensions to FCM FCM segmentation in to Grey Matter (GM), White Matter (WM) and Cerebrospinal Fluid (CSF)

6. Standard Flood-fill Flood-fill (node, target-color, replacement-color):  1. If the color of node is not equal to target-color, return.  2. Set the color of node to replacement-color.  3. Perform Flood-fill (one step to the west of node, target-color, replacement-color).      Perform Flood-fill (one step to the east of node, target-color, replacement-color).      Perform Flood-fill (one step to the north of node, target-color, replacement-color).      Perform Flood-fill (one step to the south of node, target-color, replacement-color).  4. Return. ___________________________________________________ • Theoretically sound • Realistically breaks down in a stack overflow

7. Improved Flood-Fill Flood-fill (node, target-color, replacement-color):  1. Set Q to the empty queue.  2. If the color of node is not equal to target-color, return.  3. Add node to the end of Q.  4. While Q is not empty:   5.     Set n equal to the first element of Q  6.     If the color of n is equal to target-color, set the color of n to replacement-color.  7.     Remove first element from Q  8.     If the color of the node to the west of n is target-color:  9.         Add that node to the end of Q 10.     If the color of the node to the east of n is target-color:  11.         Add that node to the end of Q 12.     If the color of the node to the north of n is target-color: 13.         Add that node to the end of Q 14.     If the color of the node to the south of n is target-color: 15.         Add that node to the end of Q 16. Return.

8. Real-World implementations • Many optimisations are possible such as: • East-West loops • Scanline Fills • Boundary condition checks • Yet most still conform to the same logical basis as the recursive algorithm

9. User driven centroid selection • Used for isolation of a specific pre-defined biologicially significant entity via a modified parallel flood-fill algorithm 3 images from a 4000x4000x4000 volumetric image, at depths at the end of the first, second, and third quarters

10. Parallelised RAM-Efficient 3D Flood-Fill • Improve Queue based Flood-Fill with East-West optimisation • Each slice of the volumetric image on the z-axis possesses it's own queue • The volumetric image is divided along the z-axis according to the number of processes available • Boundary slices (the single slices between divided segments) are left until last to prevent clashes • Only slices currently being used by a process are loaded into RAM

11. Possible example output Images from Google Body showing differing layers of anatomy of a synthetic person

12. Conclusion • There is a problem faced by many researchers dealing with large volumetric images • Current processes are insufficient to handle the increasing number and complexity of images • Parallelised 3D flood-fill with user defined centroids will assist researchers in accurately separating biologically significant entities of interest for isolated analysis