Multiple Organ detection in CT Volumes Using Random

  • Slides: 1
Download presentation
Multiple Organ detection in CT Volumes Using Random Forests Problem and Motivation Organ localization

Multiple Organ detection in CT Volumes Using Random Forests Problem and Motivation Organ localization is the process of finding the location and extent of an organ in a medical image. Daniel Donenfeld Sarfaraz Hussein dbd [email protected] edu [email protected] ucf. edu Cornell University of Central Florida Results Materials & Methods Organ Classifications: Dark Blue: Background, Yellow: Heart, Light Blue: Liver, Red: Kidney Methods Segment into supervoxels Extract hand-crafted features Classify supervoxels Smooth Data Return bounding boxes Motivation • Useful for segmentation tasks • Anomaly detection, e. g. Tumors • Fat quantification • Better medical image databases • Selectively load region of interest to reduce image size Supervoxels • Supervoxel segmentation is an over-segmentation technique • Reduce the search space • A CT scan have 50 to 100 million voxels • Can over-segment into 3000 supervoxels Location of samples around super voxel center Features Tested • • • Histogram Gradient Histogram Haar 3 D SIFT Gray Level Co-Occurrence Matrix Features Non-local patch Features Issues Small sample of a random forest, showing two trees which contribute to the classifications • There was significantly more background than relevant organ • Resolved by adding weights to the classifier to rely more on the organs Discussion 3 D Haar feature masks used State of the Art: Regression Forests for Efficient Anatomy Detection and Localization in CT Studies Results • Previous papers have used SLIC Organ Classifications: Dark Blue: Background, Yellow: Heart, Light Blue: Liver, Red: Kidney • We used a method which used SLIC superpixels, edges, and optical flow to compute supervoxels • Fast algorithm: ~6 seconds • Error approximately halved over previous state of the art atlas based registration • Acts on each voxel: each one predicts every voxel • Qualitatively better edge adhesion then SLIC supervoxels Advantages: • Significantly smaller search space • Each supervoxel has much more information than a single voxel Haar Features gave best results Haar with 20 train patients, 10 test patients Misclassification: • Near the organs, many supervoxels are misclassified • Bounding volume contains non-organ supervoxels Future Work • Use Confidence fusion to smooth classification results • Reduce the search space using hierarchical anatomical structure Limitations • Computationally expensive supervoxel and feature extraction The shape of supervoxels shown across four slices of a patient. The heart can be seen in the center, which have good adherence to the boundaries of organs. The lungs can be seen on either side in dark purple and green Acknowledgements Comparison of Haar features with different extensions Thank you to the NSF for funding the REU program for the University of Central Florida. Also, thanks to Dr. Shah, Dr. Bagci, and Dr. Lobo for overseeing the program.