Room: Karl Dean Ballroom C
Purpose: Accurate segmentation based on the image of the day is the major challenge for online adaptive replanning. Available auto-segmentation technologies cannot always generate accurate contours. This study is to investigate the feasibility of automatically correcting inaccurate contours generated from auto-segmentation using MRI texture information.
Methods: A pair of abdominal T1-weighted MRIs acquired at simulation and at a radiation therapy fraction for each of 3 patients were utilized to demonstrate the method. On each image set, initial contours generated by auto-segmentation were evaluated on a slice basis against the ground truth contours delineated by an experienced radiation oncologist. The initial contours were regarded as inaccurate if Dice Similarity Coefficient (DICE) <0.9 or Hausdorff Distance (HD)>3mm as compared with the ground truth. The inaccurate contours were re-processed using the currently proposed novel texture-based automatic contour correction method. The procedure consists of two steps: (1) calculating voxel-based GLCM-Cluster Shade texture feature map using 3Ã—3 block size inside a region-of-interest created from initial contour, and (2) incorporating the feature map into active contour algorithm as external force to find correct structure boundaries. To validate our method, DICE and HD were re-evaluated based on corrected contours.
Results: A total of 73 out of 98 initial pancreas contour slices extracted from 6 MR images sets were identified as inaccurate and selected for the contour correction. After correction, 25 out of 73 contour slices are regarded as accurate under DICE and HD criteria. Mean HD across 73 slices decreased from 10.1 mm to 6.9 mm, while mean DICE increased from 0.81 to 0.86.
Conclusion: The proposed image-texture based automatic contour correction approach can potentially improve the overall contour accuracy and the efficiency of contour modification. With further development, the method may be used as a part of segmentation process for MRI-guided online replanning.