(Sunday, 7/12/2020) [Eastern Time (GMT-4)]
Room: AAPM ePoster Library
Purpose: Existing clinical datasets typically include a mixture of well-curated contours as well as contours with unverified quality. Our goal is to provide an open-source tool to automatically segment organs-at-risk involved in the immune system regardless of contour quality. To facilitate this, we built and validated a Cascading Deep Learning Segmentation (cDLS) framework.
Methods: The cDLS framework leveraged contextual information from 20 external Computed Tomography (CT) scans with well-curated contours to mitigate the effect of learning from unverified liver clinical contours within the internal dataset. Learned representations from the external liver model were transferred to segment organs-at-risk from internal 172 CT scans of patients previously treated with thoracic RT. The scans included liver contours from treatment planning in addition to well-curated post-treatment spleen contours. The cDLS model was trained on a deep neural network architecture, DeepLabv3+ for multi-label segmentation (learning rate: 0.01; batch size: 8 images for 25 epochs for external, 50 for internal dataset) using 90% of the CT scans. The remaining 10% of the internal scans were used for quantitative evaluation of the final model vs. reference contours using Dice Similarity Coefficients (DSC), 95th percentile of Hausdorff Distances (HD95) and liver and spleen mean doses. Statistical comparison was performed using the Wilcoxon signed rank-sum test.
Results: The cDLS model reduced segmentation time per patient from about half hour of manual segmentation to 10 seconds. The liver and spleen achieved accuracies of (median DSC=(0.95(0.93-0.96), 0.94(0.91-0.94)) and HD95=(4.9mm(4.2mm-5.8mm), 3.3mm(2.9mm-3.7mm). No statistically significant difference was observed between the calculated mean doses for the auto-generated and the reference contours for both structures (p-value>=0.9).
Conclusion: The model was robust against variability in image characteristics, including the presence/absence of contrast. The accuracy was judged adequate for extracting dose-volume information for outcomes analyses. The cDLS model is distributed as part of CERR’s Model Implementations library.
Funding Support, Disclosures, and Conflict of Interest: This research is partially supported by NCI R01 CA198121.
Keywords
Computer Vision, Contour Extraction
Taxonomy
Not Applicable / None Entered.
Contact Email
