Purpose: Fast and accurate auto-segmentation is essential for MR-guided adaptive radiotherapy (MRgART). Deep learning auto-segmentation (DLAS) is not always clinically acceptable, particularly for complex abdominal organs. We have previously developed an automatic contour refinement (ACR) method based on the active contour model (ACM) that can partially correct for inaccurate contours from DLAS. This study aims to develop a DL-based ACR model to work in conjunction with ACM to improve the contour refinement.
Methods: The ACR model based on the guided DL U-Net architecture was trained and tested using the inaccurate subregion contours of bowels created by an in-house DLAS system from abdominal T2w-MRIs of 76 patients. The inaccurate contours were classified in two groups: (1) major error group with Dice similarity coefficient (DSC)<0.5 or mean distance to agreement (MDA)>8mm, and (2) minor error group with remaining contours of 0.5≤DSC<0.8 or 3mm
Results: By applying the ACM ACR, 44% (177/401) were improved to minor errors and 5% (22/401) became acceptable. Among these 177 contours with minor errors, the U-Net ACR model refined 47% (84/177) to acceptable. Overall, 36% (145/401) were turned into minor errors, and 30% (120/401) became acceptable after sequentially applying the ACM and U-Net ACR methods.
Conclusion: The obtained U-Net automatic contour refinement model substantially improves the correction of the inaccurate MRI contours of complex abdominal organs, such as bowels, thus minimizing the manual editing and accelerating segmentation process for MRgART.
Funding Support, Disclosures, and Conflict of Interest: The research was partially supported by the Medical College of Wisconsin (MCW) Cancer Center and Froedtert Hospital Foundation, the MCW Meinerz and Fotsch Foundations, and the National Cancer Institute of the National Institutes of Health under award number R01CA247960.
MRI, Segmentation, Image Processing