Click here to


Are you sure ?

Yes, do it No, cancel

Deep Learning for Head and Neck Segmentation in MR: A Tool for the MR-Guided Radiotherapy

B Anderson*, B Elgohari , C Cardenas , A Mohamed , P Yang , C Fuller , C Chung , K Brock , UT MD Anderson Cancer Center, Houston, TX


(Wednesday, 8/1/2018) 10:15 AM - 12:15 PM

Room: Karl Dean Ballroom B1

Purpose: To rapidly and robustly auto-segment normal head-and-neck anatomy on MR T1- and T2-weighted images using a fully convolutional network to enable widespread use of dose accumulation, adaptive treatment planning, and online MR-guided radiotherapy.

Methods: One hundred and thirteen patients with manually contoured parotid glands, submandibular glands, sublingual, and mandible were obtained. Image sets consisted of non-contrast-enhanced CT, and T1- and T2-weighted MR images. To date the MR images have been utilized. Ninety-three randomly selected patients were used for training, the remaining twenty patients were used for final testing. A fully-convolutional neural network with down-convolution and up-convolution steps was used for pixel-wise classification on each 2D image slice. Skip-steps were used to provide high-resolution features to the up-convolutional paths and to reduce bottle-necks in training. Transfer learning from the Visual Graphics Group (VGG) 19 pre-trained network were implemented as the first layers of the network, along with skip-layers and up-sampling for the final segmentation decision. Volume weighted cross-entropy was used due to the large imbalance between background segmented voxels, i.e. a structure with 1/10 the overall volume presented in all the images would have a weighting of 10x on its loss values.

Results: Agreement between model-created contours and physician-drawn contours was assessed using dice similarity coefficient (DSC), mean surface distance (MSD), and false negative (FN). Median agreement was found to be greater than 0.74(Dice), less than 1.4mm (MSD), and less than 0.17(FN) for parotid and mandible contours across both sequences. Submandibular and sublingual evaluation is ongoing. Contours created required <1min on a 16GM(RAM), 3.3GHz processor.

Conclusion: These results suggest that the developed fully-convolutional neural network can rapidly and accurately segment structures in the head-and-neck region on T1 or T2-weighted images. Further investigation of the neural network auto-segmentation should be conducted as a potential tool for the MR-LINAC.


MRI, CT, Linear Accelerator


Not Applicable / None Entered.

Contact Email