Program Information
Automatic Organ Contouring for Thoracic CT Images with Convolutional Encoder and Decoder Deep Learning Neural Network
D Lam*, B McClain , B Sun , T Zhang , L Santanam , S Goddu , J Bradley , S Mutic , T Zhao , Washington University School of Medicine, St.louis, MO
Presentations
TU-L-GePD-JT-3 (Tuesday, August 1, 2017) 1:15 PM - 1:45 PM Room: Joint Imaging-Therapy ePoster Theater
Purpose: Contouring is a time-consuming, error-prone process in radiotherapy. The purpose of this work is to develop a fully automated segmentation method using an artificial intelligence (AI) approach to relieve clinicians from the burden of contouring in thoracic CT images.
Methods: In this work, the contouring problem was treated as a pixel classification problem, where our algorithm predicted whether a single pixel belonged to a specific organ. A deep learning neural network architecture was created with an encoder and a decoder. The encoder created a low resolution but high level representation of the image including the organ. The decoder network was responsible for producing high-dimensional features for each pixel classification driven by dice coefficient. The method was performed on 65 patient data sets from the Kling Proton Therapy Center at the Siteman Cancer Center; 58 were used for training and 7 were used for testing. To increase the number of images for training, image augmentations such as flipping, zooming and rotation were applied. We performed auto-segmentation on 3 different organs: lung, heart and esophagus. The training was executed on NVIDIA GTX 980.
Results: It took 4 hours to train the whole network end to end with the footprint of GPU memory about 3GB. The result obtained on the test sets with no prior knowledge of existing contours demonstrated qualitatively consistent contours compared to the physician contouring work. Segmentation accuracy measured in term of Dice coefficient for lung, heart and esophagus were 95.65%, 85.27% and 65.67% respectively.
Conclusion: AI promises useful for fully automatic contouring. In this work, with the advantage of deep learning in pixel-wise classification, encouraging results were achieved. Future work will concentrate on increasing the robustness and accuracy of the segmentation algorithm and the number of clinical sites by collecting more data and designing deeper neural network.
Funding Support, Disclosures, and Conflict of Interest: Mevion Research Grant
Contact Email: