Initial Progress Report Daniel Donenfeld
Contents Goal Approach Timeline Super-Voxels Progress
Full CT with Bounding Volumes Project Goal Generate organ Bounding Boxes Detect Heart, Kidney, Liver Full CT scan Full CT with Bounding Volumes
Approach - Training Training Train Random Forest on super-voxels Train organ displacement
Approach - Testing Reduce Search space with location priors Segment into supervoxels Run each super-voxel through random forest Improve estimates using inter-organ displacement Return bounding boxes
Initial Timeline Research supervoxel implementations and uses in medical imaging Over-segment scans with supervoxel Train and test random forest on supervoxels Extract inter-organ displacement Integrate organ displacement data
Super-Voxel Over-Segmentation Pixels are an artifact of digital sampling Useful to extract more relevant regions from image Reduce computation One of the best superpixel algorithms is SLIC Also extendable to supervoxels
Super-Voxel using SLIC Initialize cluster centers
Super-Voxel using SLIC Move each cluster center to the lowest nearby gradient position For each cluster center Do k-means clustering in a small region surrounding the cluster Repeat until error threshold is reached
Why SLIC? CRCV lab researches motion in videos using supervoxels Algorithms would not work effectively for CT images Ex: Spatio-Temporal Object Detection Proposals, uses SLIC on image slices, then optical flow between slices Ex: LIBSVX, Uses color information
Why SLIC? Literature reveals its one of the prefered methods for segmenting medical images Useful for CT scans Graph cut based Automatic Prostate Segmentation Near-optimal Keypoint Sampling for Fast Pathological Lung Segmentation Edge- and detail-preserving sparse image representations for deformable registration of chest MRI and CT volumes
Progress Find SLIC implementation to generate supervoxels Lots of implementations of SLIC for superpixel (in general and matlab) Fewer implementations of SLIC for supervoxel
Progress Implementations exist in python and C++ Can call python from matlab Have used Scikit (python) implementation of SLIC problems interfacing, cannot enforce connectivity Plan - write small wrapper to easily call C++ implementation Use results in matlab
Using Python