Program Information
Level Set Based Segmentation with a Dynamic Shape Prior
W Liu1*, D Ruan1,2,3 , (1) Department of Bioengineering, University of California, Los Angeles, CA, (2) Department of Radiation Oncology, University of California, Los Angeles, CA, (3) Department of Biomedical Physics, University of California, Los Angeles, CA
Presentations
TU-F-BRF-5 Tuesday 4:30PM - 6:00PM Room: Ballroom FPurpose: To perform accurate segmentation on low-SNR MR images subject to motion artifacts and signal voids.
Methods: We have developed a novel level set based segmentation method that incorporates a dynamic shape prior. In contrast to conventional shape prior models that are either based on a single template or statistical models, we modeled the shape prior as sparse linear combinations of templates in a shape library. The proposed method was applied to segment real-time kidney geometry from an abdominal MR series acquired under respiration and compared with the Chan-Vese approach. Synthetic occlusions were further introduced to assess performance robustness. We compared the segmentation results with the ground truth contours and evaluated the segmentation accuracy based on the Dice similarity coefficient.
Results: The proposed method successfully segmented all kidneys in the testing series. The corresponding shape priors were reasonably reconstructed from a small number of shape templates in the library, as expected. On testing dataset without occlusion, the proposed method achieved mean DSC of 0.97, which was more accurate than the Chan-Vese approach with a DSC of 0.93. When tested on dataset with synthetic occlusions, the proposed method made robust inference on the occlusion sites and achieved mean DSC of 0.95, compared to 0.76 from the Chan-Vese method, demonstrating the advantage of incorporating the proposed shape prior.
Conclusion: We have developed a novel level set based segmentation method with a novel regularization to incorporate a dynamic shape prior. The shape prior is dynamically updated during the segmentation process as a sparse linear combination of templates from a shape library. Preliminary results have demonstrated the ability of the proposed method in improving segmentation accuracy, especially when noises and/or signal voids are present. Future work will consider dynamically update and evolve the shape library to encode shape priors as segmentations are being performed.
Contact Email: