Program Information
Deep Learning Contouring of Thoracic Organs At Risk
D Peressutti1 , P Aljabar1 , J van Soest2 , T Lustberg2 , J van der Stoep2 , A Dekker2 , W van Elmpt2 , M Gooding1*, (1) Mirada Medical Ltd, Science and Medical Technology, Oxford, UK, (2) Maastricht University Medical Centre, Department of Radiation Oncology MAASTRO - GROW School for Oncology Developmental Biology, Maastricht, NT
Presentations
TU-FG-605-7 (Tuesday, August 1, 2017) 1:45 PM - 3:45 PM Room: 605
Purpose: To evaluate the potential of deep learning for automatic contouring of organs at risk in the thorax.
Methods: Deep learning methods represent a subset of artificial intelligence algorithms based on complex artificial neural networks, with broad application potential. In this investigation, a Deep Learning Contouring (DLC) method was evaluated for automatic contouring of organs at risk (OARs) in CT images of the thorax. Evaluation was performed on the lungs, spinal canal, heart, mediastinum envelope and esophagus. A set of 572 clinical cases was randomly divided into a training set (450), cross-validation set (56) and test set (66). Clinical cases were acquired from a single institution. The performance of DLC was compared against an atlas-based auto-segmentation (ABAS) method, which may be considered the current state-of-the-art for automatic contouring in radiotherapy. Multi-atlas ABAS employed a separate set of 20 atlases, from the same institution, to contour the test images. The Dice Similarity Coefficient (DSC), average distance (AD) and root-mean square distance (RMSD) were computed between both sets of automatically generated contours and the manual clinical contours.
Results: Results showed that DLC outperformed ABAS for the majority of OARs, with the exception of the heart where performance of the two methods was similar. However, DLC was found to have larger outliers in a few atypical cases, where ABAS also performed poorly. Median DSC scores for DLC vs ABAS for right lung 0.994 vs 0.981 (p<0.001), left lung 0.994 vs 0.979 (p<0.001), spinal canal 0.884 vs 0.818 (p<0.001), esophagus 0.741 vs 0.478 (p<0.001), heart 0.899 vs 0.902 (p=0.44) and mediastinum envelope 0.938 vs .923 (p=0.057).
Conclusion: Deep Learning Contouring (DLC) has the potential to significantly outperform ABAS for automatic contouring of OARS in CT images of the thorax. Consequently, DLC may lead to a further reduction of editing time required after auto-contouring.
Funding Support, Disclosures, and Conflict of Interest: This research was funded via InnovateUK Grant 600277 as part of Eurostars Grant E!9297. DP, PA, MG are employees of Mirada Medical Ltd.
Contact Email: