Purpose: Contour interpolation is an important tool to expedite manual segmentation of anatomical structures by allowing users to contour on discontinuous slices and fill-in the gaps. The conventional interpolation tool in clinical TPS is based on distance map interpolation and has difficulties in generating satisfactory results when the existing manual contours differ dramatically. This is especially apparent near the superior and inferior borders of organs, and for the gastrointestinal structures such as bowel. In this study, we developed a deep learning model to provide improved contour interpolation robustness and accuracy for these historically difficult cases.
Methods: The deep contour interpolator was developed and trained using selected 5882 datasets containing images of multiple organs (lung, liver, esophagus, heart). Preprocessing steps include cropping, image resizing and intensity normalizing. A 3D UNet model was utilized and its softmax layer and segmentation layer were replaced by a regression layer. The input size is 128×128×8 in 2 channels. The first channel is the top and bottom slices of a 5-slice patch and three middles slices duplicated. The second channel is the related masks, empty mask for middle slices. The outputs are the predicted organ masks for the 3 middle slices.
Results: The dice coefficients and DTA (distance-to-agreement) values were computed to evaluate the performance. For 1117 independent testing cases, the dice score was 0.90±0.13 for the difficult cases, comparing to 0.85±0.17 by the conventional method. The max and 95th percentile DTA values, were 7.6±13.2 and 4.5±11 mm, comparing to 8.7±14.8 and 5.2±5.2 mm by the conventional method. A t-test was conducted that confirmed that our model statistically significantly out-performed the conventional method (p<0.05).
Conclusion: These results suggest that the deep interpolator may produce more robust and accurate interpolation results than the conventional contour interpolation method for historically difficult cases.