Corrective-annotation auto-completion enables faster organ contouring
PD-0065
Abstract
Corrective-annotation auto-completion enables faster organ contouring
Authors: Abraham Smith1,2, Jens Petersen1,2, Isak Wahlstedt2,5, Signe Lenora Risumlund2, Mette Van Overeem Felter3, Vibeke Nordmark Hansen2, Ivan R Vogelius2,4
1University of Copenhagen, Computer Science, Copenhagen, Denmark; 2Rigshospitalet, Department of Oncology, Copenhagen, Denmark; 3Herlev and Gentofte Hospital, Department of Oncology, Copenhagen, Denmark; 4University of Copenhagen, Department of Health and Medical Sciences, Copenhagen, Denmark; 5Technical University of Denmark, Department of Health Technology, Kongens Lyngby, Denmark
Show Affiliations
Hide Affiliations
Purpose or Objective
Corrective-annotation methods have been shown to continually improve segmentation performance and consistency for routine delineation tasks. Correcting auto-generated contours may still be tedious as error regions often span multiple axial slices, requiring similar manual corrections to each slice. We propose to use the ongoing corrective annotation of the current scan as input to the model to refine predictions in subsequent slices, generating real time auto-contour updates as delineation progresses.
We evaluate the performance of our contour auto-complete method by comparing it to a baseline method where all corrections are assigned manually without updates to the model prediction during delineation (full-correction). We hypothesize that the auto-complete method will result in reductions in delineation time in comparison to the baseline method whilst maintaining high accuracy.
Material and Methods
Both corrective annotation methods (auto-complete and full-correction) were implemented as variations of the open-source RootPainter3D [1] deep learning auto-contouring software, which uses corrective-annotation in training to improve auto-contouring accuracy.
We used a dataset of MRI scans which included multiple scans from 31 different patients with liver metastases that had been referred to SBRT. 177 scans were delineated in the same order using both methods. 12 images from the 177 were also delineated by a trained clinician using the MRIdian planning system to enable consistency to be checked between the completed corrective delineations and standard clinical delineations.
The corrective delineation procedure consisted of 3 annotation sessions, with models left training overnight, on the newly expanded dataset of annotations, between each session. The sessions included 20, 60, and 97 images, respectively. Delineation time was automatically recorded, including time to both review and correct identified mistakes.
Results
Figure 1: Delineation duration for each scan in order of completion time for both the auto-complete and full-correction methods. Auto-complete starts out slower than full-correction, but becomes significantly faster as more images are delineated. For comparison, manual liver delineation has been reported as taking between 4 and 8 minutes.
Figure 2: The mean dice with expert clinician delineations for the completed contours for both methods was similar to inter-annotator variation (~0.94), indicating high accuracy and suitability for clinical use.
Conclusion
During the first annotation session, the auto-complete method hampered contouring performance, as poor contour auto-completions increased the delineation workload but with sufficient training data and time, corrective annotation with auto-completion provided some reductions in contouring time, whilst maintaining high delineation accuracy.
[1] https://arxiv.org/abs/2106.11942