1. Multiple Organ detection in CT Volumes Using Random Forests - Week 6
Daniel Donenfeld

2. Work This week Tested features on more samples
3D SIFT is not discriminative enough for our purposes.
Began work on using distance priors

3. Results - Histogram Features
Percent Liver Correct vs Number of Trees
Percent Heart Correct vs Number of Trees
Number of Trees
Number of Trees

4. Results - Histogram Features
Percent Left Kidney Correct vs Number of Trees
Percent Right Kidney Correct vs Number of Trees
Number of Trees
Number of Trees

5. Results - 3DSIFT Features
Classified every supervoxel as background
Verified the classifier by training SVM on 3D SIFT features with same results
Can test again after removing normalization steps
Currently worse results than histogram features
Have not yet re-tested 3D HOG

6. Distance Priors There are 50,000,000-100,000,000 voxels
Using supervoxels search space is reduced to 3000 supervoxels
Want to reduce search space further

7. Distance Priors We have prior information about the location of organs
Human Anatomy dictates general location of organs within a person
Just search the area where the organ is going to be located

8. Registering Images People are all different sizes
Register images to single space
Then there is a common set of coordinates

9. Image Registration and Atlas
Used to normalize scans of a single patient
Progression of COPD in the lungs1
Used global affine transform and B-splines
Mean sum of squared distances similarity
Both are tailored for intra-patient computations
1. Gorbunova, V., Lo, P., Ashraf, H., Dirksen, A., Nielsen, M., & Bruijne, M. (n.d.). Weight Preserving Image Registration for Monitoring Disease Progression in Lung CT. Lecture Notes in Computer Science Medical Image Computing and Computer-Assisted Intervention – MICCAI 2008,

10. Image Registration and Atlas
Whole Heart Segmentation2
Locally Affine Registration Method
Global affine is not accurate enough for segmentation
Non-rigid transform could affect topology
Abdominal Segmentation3
Mutual information similarity
Warping transform - TPS
Use of multiple atlases and confidence fusion4
2. Zhuang, X., Rhode, K., Arridge, S., Razavi, R., Hill, D., Hawkes, D., & Ourselin, S. (n.d.). An Atlas-Based Segmentation Propagation Framework Using Locally Affine Registration – Application to Automatic Whole Heart Segmentation. Lecture Notes in Computer Science Medical Image Computing and Computer-Assisted Intervention – MICCAI 2008,
3. Park, H., Bland, P., & Meyer, C. (n.d.). Construction of an abdominal probabilistic atlas and its application in segmentation. IEEE Transactions on Medical Imaging IEEE Trans. Med. Imaging,
4. Acosta, O., Dowling, J., Drean, G., Simon, A., Crevoisier, R., & Haigron, P. (2013). Multi-Atlas-Based Segmentation of Pelvic Structures from CT Scans for Planning in Prostate Cancer Radiotherapy. Abdomen and Thoracic Imaging,
Mutual information - MI(F, M ) = H(F) + H(M ) - H(F, M)

11. Image Registration - Affine
Before Registration
After Registration

12. Image Registration