Program Information
Prediction of Lung Tumor to Normal Tissue Interface Type Using Deep Learning to Support Target Definition
R N Mahon1*, E Weiss1 , K Karki2 , G D Hugo1 , (1) Virginia Commonwealth University,Richmond, VA, (2) Virginia Commonwealth University, Brookfield, WI
Presentations
TU-L-GePD-J(B)-4 (Tuesday, August 1, 2017) 1:15 PM - 1:45 PM Room: Joint Imaging-Therapy ePoster Lounge - B
Purpose: To develop an algorithm to infer the lung tumor/normal tissue interface type using a convolutional neural network to support lung tumor delineation and margin selection.
Methods: A consensus tumor contour was determined for each of nine lung cancer patients by computing the median contour from seven individual physician delineations on PET/CT images. Lung tumor/normal tissue interface classes were then identified by an experienced radiation oncologist on the nine median contours. Labeled CT image patches (12.5mm x 12.5mm x 12.5mm) were extracted to create a training dataset having seven different labels: lung/tumor, hilum/tumor, chest wall/tumor, atelectasis/tumor, aorta/tumor, mediastinum/tumor, and not interface (which describes voxels not included in any interface contour). An eight-layer, two-stage 3D volumetric convolution neural network (CNN) was trained de novo on the dataset to predict the interface type. The first stage was trained on a randomly sampled, across-subjects balanced dataset containing over 13,000 patches sampled equally from each interface type. The second phase retained the weights from the first phase and re-trained on 72,000 patches containing samples from each interface type proportional to the observed frequency. Testing was performed on an unseen dataset of 38,000 samples and classification accuracy was measured.
Results: The accuracy of the network prediction was 65% for top one guess in the first phase of training and 70% following the second phase. The individual classes had the following error rates: atelectasis/tumor 18.7%, not interface 16.8%, hilum/tumor 13.4%, lung/tumor 10.9%, mediastinum/tumor 8.7%, chest wall/tumor 3.8% and aorta/tumor 0.6%.
Conclusion: The network was able to distinguish between different tumor interface types, which could be used, for example, to assist physician delineation by recommending an interface type and estimating expected interface-specific inter-observer uncertainty in a region.
Funding Support, Disclosures, and Conflict of Interest: GDH and EW receive research support from Philips Healthcare and Varian Medical Systems and have licensed technology to Varian Medical Systems. EW receives royalties from UpToDate. RNM and KK have not no conflicts.
Contact Email: