(Tuesday, 7/31/2018) 9:30 AM - 10:00 AM
Room: Exhibit Hall | Forum 5
Purpose: The spinal column is an important dynamic structure in the human body. Accurate segmentation of individual vertebrate can have important implications in image guided radiotherapy. However, the multitude of individual, closely spaced moving vertebrae, heterogeneous density and complex morphology has made it impractical to manually, and challenging to automatically, segment them using conventional graph methods. This work aims to demonstrate a novel method of segmenting individual vertebrae using deep learning and generative adversarial networks.
Methods: A set of 10 training and 5 testing CT images with corresponding ground truth labels for the thoracic and lumbar vertebrae were acquired from the 2014 MICCAI 2014 Spine Challenge. A generative adversarial network was trained using sagittal 2D projections (each reduced to 256x256 for computational efficiency) of the CT training set. This is a network which is trained by pitting two networks against one another: one tries to generate a segmentation given an input image, while the other attempts to discern whether a segmentation is real or synthesized. The resulting segmentations were created by thresholding the output. These were compared to the ground truth using the Dice Coefficient, False Positive Rate, and False Negative Rate.
Results: The average Dice coefficient was 0.87 (range: 0.81-0.89) across the 5 testing images. The average False Positive Rate was 0.25 (range: 0.19-0.36). The average False Negative Rate was 0.05 (range: 0.02-0.08). This shows that the automatic segmentation can sometimes be too sensitive. Qualitatively, it was noted that the more superior vertebrae were more difficult to segment, possibly due to narrower intervertebral spacing and downsampled CT images, which will be remedied by network training using fully-sampled imaged.
Conclusion: We have shown that automatic segmentation of the individual vertebrae can be accomplished using a deep neural network. Automated segmentation showed good quantitative agreement with manual segmentation.
Funding Support, Disclosures, and Conflict of Interest: NIH U19AI067769 DE-SC0017687 NIH R21CA228160 DE-SC0017057 NIH R44CA183390 NIH R43CA183390 NIH R01CA188300
Keywords
Segmentation, Computer Vision, CT
Taxonomy
Contact Email
