Automatic Segmentation Algorithms Improve the Efficiency and Inter-Observer Agreement of Gross Target Volumes Contours
E Czeizler1, J Pajula2, H Polonen3, K Antila4, K Lehtio5, J Lehto6, A Tiulpin7, A Maslowski8, M Hakala9, E Kuusela10, K Bush11*, (1) Varian Medical Systems Finland, Helsinki, FI, (2) VTT Technical Research Center of Finland, Tampere, ,FI, (3) VTT Technical Research Center of Finland, Tampere, FI, (4) VTT Technical Research Center of Finland, Tampere, FI, (5) Oulu University Hospital, Department of Oncology and Radiotherapy, Oulu, FI, (6) Oulu University Hospital, Department of Oncology and Radiotherapy, Oulu, ,FI, (7) Ailean Technologies Oy, FI, (8) Varian Medical Systems, Palo Alto, CA, (9) Varian Medical Systems Finland, Helsinki, FI, (10) Varian Medical Systems Finland, Helsinki, FI, (11) Stanford University, Stanford, CA
Presentations
PO-GePV-M-345 (Sunday, 7/10/2022) [Eastern Time (GMT-4)]
ePoster Forums
Purpose: To decrease both contouring time for gross tumor volumes (GTVs) and inter-observer variability for these contours.
Methods: We propose to use a 3D-Convolutional Neural Network model to automatically segment the GTV contours starting from CT and PET images. The dataset used in this project was the Head-Neck-PET-CT dataset openly available in TCIA. First, we used OAR segmentation models to generate the contours for Brain, Brainstem, Eyes and Spinal Cord, which defined the 3D-volume of interest for our GTV segmentation model. Additionally, the maximum brain intensity was used as a stable scaling factor to normalize PET data. The impact of using model predictions when delineating GTV contours was evaluated by two clinicians. First, they manually drew the GTV contours for a set of 20 patients. Then, they redrew these structures by correcting the contours generated by our model. We then compared both the time needed in the two approaches as well as the quality of the resulting contours.
Results: There was a substantial productivity increase (20-40%) when clinicians corrected the segmentations produced by our model. We then took a deeper look at the quality of these structures and focused on a comparison between the primary GTV contours (pGTVs). The average Dice scores between the pGTVs generated by the two clinicians increased from 0.79 to 0.836 when they both corrected the automatically generated contours instead of manually delineating them. Moreover, the latter similarity score is larger than the average Dice scores between the manually delineated contours and the corrected ones for each of the two clinicians, which were 0.822 and respectively 0.804.
Conclusion: We observed a clear improvement in the productivity of the segmentation task when the clinicians started from the automatically generated contours. There was also a clear decrease in the inter-observer variability of these contours.
Funding Support, Disclosures, and Conflict of Interest: This study was partially funded by Business Finland and Varian Medical Systems. Several authors are employed by Varian Medical Systems.
Keywords
Treatment Planning, Segmentation, Observer Performance
Taxonomy
IM/TH- Image Analysis (Single Modality or Multi-Modality): Machine learning
Contact Email