(Tuesday, 7/16/2019) 7:30 AM - 9:30 AM
Room: Stars at Night Ballroom 2-3
Purpose: Image segmentation is necessary for treatment planning but manual delineation of the organs-at-risks (OARs) is time-consuming. This limitation has become a major bottleneck in abdominal adaptive MR guided Radiotherapy (MRgRT) where the time for planning is highly constrained. Current solution based on deformable registration for automated segmentation produces unsatisfactory results that need extensive manual editing. We propose a more robust solution using deep neural networks for the segmentation of low field 3D abdominal MR images.
Methods: 0.35T TrueFisp MR images were obtained for 42 abdominal patients underwent radiotherapy with an MR guided radiotherapy system. Manual segmentation for treatment planning were used to train the networks and evaluate the neural network segmentation performance. To facilitate gradient propagation through the network, ease the training of the segmentation network, and improve the network performance, a densely connected fully convolution neural network was employed for voxel prediction. Furthermore, multi-scale feature fusion and deep supervision technique were utilized to integrate fine details for accurate segmentation and speed up the network convergence. The 42 images were divided into 30 and 12 for training and validation, respectively. The stomach, duodenum, liver, cord, and kidneys (both left and right) were segmented from the abdominal MR images by the proposed algorithm.
Results: Compared with the manual segmentation, the following average Dice’s indices were achieved: 0.85 (stomach), 0.70 (duodenum), 0.92 (liver), 0.75 (cord), 0.93 (kidney). The segmentation accuracy of duodenum is affected by the ambiguous boundary between the duodenum and the stomach, which may be inconsequential in planning.
Conclusion: A fully automated segmentation method for multi-organ segmentation from low field abdominal MR images was developed and evaluated. The study demonstrates that the proposed algorithm can provide fast, accurate and robust segmentation for abdominal adaptive MRgRT without relying on the questionable deformable registration.
Funding Support, Disclosures, and Conflict of Interest: This work was supported by the CSC Chinese Government Scholarship, NIH Grants (R01CA188300).
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
