Multi-Scale Deep Learning CT Liver and Spleen Segmentation Network for Radiation-Induced Toxicity Outcomes Analysis
R Haq*, I Onochie, A Apte, M Thor, J Deasy, Memorial Sloan-Kettering Cancer Center, New York, NY
Presentations
PO-GePV-M-243 (Sunday, 7/25/2021) [Eastern Time (GMT-4)]
Purpose: Dose toxicity to immune-related organs-at-risk during radiation therapy (RT) may contribute to lymphopenia and radiation-induced liver disease. Our goal is to provide an open-source tool to automatically segment the liver and spleen from clinical contours to enable maximal tissue sparing and perform clinical outcomes analysis. To facilitate this, we built and validated a Deep Learning Segmentation (DLS) framework for consistent and robust auto-segmentations.
Methods: The DLS framework leveraged contextual information from 182 contrast and non-contrast enhanced Computed Tomography (CT) scans with well-curated clinical contours of patients previously treated with thoracic/abdominal RT. The DLS model was trained on a nested deep neural network architecture, Unet++, for multi-label segmentation (learning rate: 0.1; batch size: 8 images for 70 epochs) using 90% of the CT scans. The multi-scale, deep-supervision enabled network learned representations from dense skip-connections on entire CT body scans. The remaining 10% of the 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 the mean dose. Statistical comparison was performed using the Wilcoxon signed-rank test and Spearman’s correlation coefficient tests.
Results: The DLS model reduced segmentation time per patient from about half an hour of manual segmentation to 40 seconds. The liver and spleen achieved accuracy of (median DSC=(0.94(0.93-0.95), 0.92(0.89-0.93)) and HD95=(7.5mm(5.3mm-8.4mm), 3.7mm(3.3mm-6.2mm). No statistically significant difference was observed between the calculated mean dose for the auto-generated and the expert contours for both structures (p-value>=0.4).
Conclusion: The model was robust against variability in image characteristics and large image background, including the presence/absence of contrast. The segmentation accuracy is considered adequate for extracting dose-volume histograms to use in outcomes analyses. The abdominal DLS model is distributed as a part of CERR’s Model Implementations library.
Funding Support, Disclosures, and Conflict of Interest: This research is partially supported by NCI R01 CA198121.
ePosters
Keywords
Segmentation, Computer Vision, CT
Taxonomy
Contact Email