Click here to


Are you sure ?

Yes, do it No, cancel

Standardizing Structure Labels in Head-And-Neck Region Using Non-Local 3D-ResNet

Q Yang1,2*, T Rozario1, A Barragan Montero1, M Joo1, R McBeth1, H Chao1,2, S Jiang1, (1) Medical Artificial Intelligence and Automation (MAIA) Laboratory, Department of Radiation Oncology, University of Texas Southwestern Medical Center, Dallas, Texas, USA, (2) School of Data and Computer Science, Sun Yat-sen University, Guangzhou, Guangdong, P.R.China


(Tuesday, 7/16/2019) 11:00 AM - 12:15 PM

Room: Stars at Night Ballroom 2-3

Purpose: To automatically standardize nomenclatures that applied to organs at risk (OARs) in the head-and-neck (HN) region of CT scans.

Methods: There are two models, a 2-channel model and a 3-channel model. For the 2-channel model, the first input channel is 12 slices of the cropped CT images and the second channel is the mask of the OAR to be labeled in the same slices. For the 3-channel model, we added a third channel that is composed of a weighted sum of all other OAR masks in the same 12 slices. We trained both models with three kinds of data augmentation: random affine augmentation, over-sampling in the minority categories, and multi-scale down-sampling in preprocessing. Finally, we added a non-local block to the end of each residual block in res3 and res4 and built the non-local 3D-ResNet architecture.

Results: We divided the open source dataset HN-PET-CT (298 patients) into three datasets for training, validation, and testing. This dataset has 28 OAR categories in the HN region. To demonstrate the generalization ability of the models, we tested the trained models on other two datasets: PDDCA (48 patients with 9 OAR categories) and our own HN datasets (406 patients with 28 OAR categories). The final models gain 99.46% and 97.06% average recall rate on PDDCA and our own HN datasets, respectively.

Conclusion: We proposed a method to re-label the 28 categories of OARs in the HN region and train models with an extremely imbalanced dataset. The final models have remarkable performance on categorizing 28 HN OARs, and they could be extended for all OARs in the whole body in the future work.


Classifier Design, Computer Vision, CT


Not Applicable / None Entered.

Contact Email