Purpose: Fast and accurate auto-segmentation on daily image is essential for MR-guided adaptive radiotherapy (MRgART). However, the state-of-the-art auto-segmentation based on deep learning (e.g., deep convolutional neural networks, DCNN) has a limited success particularly for complex structures, e.g., bowels. To address this issue, we develop a method to quickly and automatically correct for unacceptable auto-segmented contours for MRgART.
Methods: An improved active contour model (ACM) based on level set was implemented to utilize the probability maps from standard DCNN models to set per-pixel parameters and initialize contour evolution. The ACM method was tested on the DCNN auto-segmented bowel contours (initial contours) for 46 abdominal T2w-MRI sets. All MRI images were pre-processed with bias correction and contrast enhancement. For each image slice, bowel region was divided into multiple subregions with each including a bowel loop. The ACM contour correction was performed on the initial contour in each subregion. The correction performance was measured by comparing the Dice Similarity Coefficients (DSC) for the corrected contours with those for the initial contours. Manually delineated contours were used as ground truth.
Results: The DSC for the ACM-corrected contours was increased compared to those for the initial contours. For bowel contours with initial DSC<0.5, the mean DSC was increased from 0.307 to 0.574 after correction, with 16% of the corrected contours reaching DCS≥0.8. For contours with DSC=0.5~0.8, the mean DSC was increased from 0.694 to 0.773 after correction, with DCS≥0.8 in 47% of cases. The execution time of the correction on a subregion was less than 10 seconds.
Conclusion: A fully-automated contour correction method based on ACM was performed to quickly correct for inaccurate deep learning auto-segmentation and its effectiveness was demonstrated for complex anatomy, e.g., bowels. The method can be integrated into the segmentation step to improve and accelerate the process of MRgART.
Funding Support, Disclosures, and Conflict of Interest: Funding support from NIH R01CA247960.