Exhibit Hall | Forum 8
Purpose: Deep learning auto-segmentation (DLAS) developed for MRI-guided adaptive radiotherapy (MRgART) has very limited success for complex abdominal organs, requiring labor-intensive and time-consuming editing after DLAS. This work aims to develop a DL method based on Dense UNet to automatically correct the inaccurate DLAS contours, thus minimizing the subsequent editing time.
Methods: Dense UNet, a DL algorithm that combines UNet with dense blocks, was trained to create automatic contour correction (ACC) models. The data used for training, validation, and testing included (1) 76 abdominal MRI sets with ground truth contours and contours created using an in-house DLAS tool and (2) 650 MRI sets created by augmenting the positions, shapes, and sizes of the organs and the ground truth contours and DLAS contours. For this proof of principle study, two models were trained for the stomach using two categories of data: DLAS contours requiring major or minor edits, which were determined using a contour classification model based on seven quantitative metrics that was developed in a separate study. The obtained model for minor edits was tested on contours directly from the DLAS. The model for major edits was applied sequentially on contours first corrected by a previously developed ACC method based on active contour method (ACM). The performance of the Dense UNet models was quantified by comparing the quality of the contours before and after the ACC.
Results: After correction with Dense UNet, 33% of DLAS contours needing minor edits became acceptable. After Dense Unet with ACM pre-correction, 32% of slices with DLAS contours needing major edits became acceptable and 43% shifted to requiring minor edits.
Conclusion: The Dense UNet ACC model is a useful tool for automatically improving the quality of contours from auto-segmentation on abdominal MRI and can substantially reduce contour editing time for complex anatomy during MRgART.
Image Guidance, MRI, Segmentation