Multiple Organ detection in CT Volumes Using Random Forests - Week 7 Daniel Donenfeld
Results All new features were not discriminative enough due to small ratio of positive to negative data Tried training and testing on only ground truth positive examples Tested features on patches first to verify that features were working Wrote 3D haar-like code Tested SPOT on patient
Large Amount of Background There are 3000 supervoxels in each patient Around 3 percent of those are organs Huge amount of background supervoxels trained classifier to recognize everything as background
3D Haar-Like features Paper on using 3D haar-like features: Cui, X., Liu, Y., Shan, S., Chen, X., & Gao, W. 3D Haar-Like Features for Pedestrian Detection. Multimedia and Expo, 2007 IEEE International Conference on. Described extending haar to 3D Implemented the method they proposed with filters on right
Results No Background- Histogram
Results No Background- Gradient
Results No Background - GLCM
Random Sample BG - Histogram Randomly sampled 1000 background supervoxels
Next Week Plans Investigate using SPOT and confidence fusion to improve results from supervoxels Find optimal sampling amount of background Continue tuning feature vectors