Purpose: Accurate and robust auto-segmentation of highly deformable organs, e.g. stomach, remains an outstanding problem due to these organs’ frequent and large anatomical variances. We propose a machine-assisted interpolation (MAI) that uses prior information, in the form of sparse manual delineations, to auto-segment the stomach in low field magnetic resonance images (MRI).
Methods: Images from 96 patients undergoing 0.35T MRI-guided abdominal radiotherapy were collected. For each patient volume, the manual delineation of the stomach was extracted from every 8ᵗʰ slice. These manually drawn contours were first interpolated to obtain a rough segmentation mask of the stomach. A 64x64 pixel patch-based convolutional neural network (CNN) was first used to localize the position of the organ’s boundary on each slice within a 5-pixel wide road. This boundary prediction was then used as an anatomical constraint for a second CNN, which output the final stomach segmentation. Both networks used a Dense-UNet architecture and were trained, validated, and tested on 79, 7, and 10 patients respectively. Algorithm performance was compared against linear interpolation and de novo auto-segmentation using a Dense-UNet trained on the same dataset.
Results: MAI resulted in a mean 50ᵗʰ Hausdorff Distance (50HD) of 1.6 ± 1.1 mm and mean Dice Similarity Coefficient (DSC) of 0.91 ± 0.07 with 6% of slices with DSC<0.8. De novo auto-segmentation resulted in 50HD of 4.7 ± 11.8 mm and DSC of 0.80 ± 0.24 with 22% of slices having DSC<0.8. Linear interpolation resulted in 50HD of 2.7 ± 2.3 mm and DSC of 0.86 ± 0.12 with 26% of slices having DSC<0.8. Differences in 50HD, DSC, and fraction of slices with DSC<0.8 were all statistically significant between MAI and comparison methods (p<0.001).
Conclusion: The proposed machine-assisted interpolation algorithm significantly outperformed de novo segmentation and linear interpolation in terms of accuracy and robustness for stomach segmentation.
Funding Support, Disclosures, and Conflict of Interest: Research reported in this abstract was supported by the Agency for Healthcare Research and Quality under award number 1R01HS026486.
Segmentation, Image Analysis, Low-field MRI