Click here to


Are you sure ?

Yes, do it No, cancel

Development and Validation of Deep Learning Segmentation Network for Cardio-Pulmonary Substructure Segmentation

R Haq*, A Hotca-Cho , A Apte , A Rimner , J Deasy , M Thor , Memorial Sloan Kettering Cancer Center, New York, NY


(Tuesday, 7/16/2019) 7:30 AM - 9:30 AM

Room: Stars at Night Ballroom 2-3

Purpose: The results of Trial RTOG 0617 indicate that doses to some cardio-vascular sub-structures may be critical factors in the observed early mortality following radiotherapy (RT). Our goal is to provide an open-source tool to automatically segment cardio-vascular substructures for consistent outcomes analyses. To facilitate this, we built and validated a Deep Learning Segmentation (DLS) framework for accurate auto-segmentation of cardio-pulmonary substructures.

Methods: The DLS framework utilized a deep neural network architecture, DeepLabv3+, to segment eight cardio-pulmonary substructures (Aorta, Left Atrium (LA), Right Atrium (RA), Left Ventricle (LV), Right Ventricle (RV), Superior Vena Cava (SVC), Inferior Vena Cava (IVC) and Pulmonary Artery (PA)) from Computed Tomography (CT) scans of 241 patients previously treated with thoracic RT. The scans included a wide range of scanning parameters and absence/presence of contrast-enhancement. The DLS model was trained for multi-label segmentation (learning rate: 0.01; batch size: 8 images for 50 epochs) using 80% of the CT scans and reviewed on 10%. The remaining 10% of the scans were used for quantitative evaluation of the final model vs. expert contours using Dice Similarity Coefficients (DSC) and 95th percentile of Hausdorff Distances (HD95). A sample of the results was reviewed by an expert.

Results: The DLS model reduced segmentation time per patient from about one hour for manual segmentation to 17 seconds. Quantitatively, the highest accuracy was observed for LV (median DSC=(0.92(0.91-0.93)) and HD95=(4.3mm(3.8mm-5.5mm)). The median DSC for the remaining structures was 0.80-0.92. The expert judged that, on average, 83% of slices were equivalent to state-of-the-art manual contouring.

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 histogram information for outcomes analyses. The 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.


Dose Response, Segmentation, CT


IM- CT: Segmentation

Contact Email