Room: Exhibit Hall | Forum 2
Purpose: To develop an automatic organs-at-risk (OAR) name classifier to identify and correct inconsistently and incorrectly labelled OARs in retrospective data to facilitate big data studies.
Methods: We developed a 3D convolutional neural network (CNN)-based architecture to classify 19 normal structures in the head-and-neck region (esophagus, parotids, larynx, brainstem, cord, cochleae, eyes, brain, oral cavity, optic nerves, optic chiasm, submandibular, and lenses), and named it TG263-Net. TG263-Net, inspired by the AAPM report on proper nomenclature, takes a two-channel 3D image as an input, in which the first channel provides the network with a binary mask of the normal structure and the second channel provides a 3D patch of the CT scan. For data augmentation, 8 extra images were generated per image by translating the center-of-mass by 10mm in the x-, y-, and z-directions. Our architecture is loosely based on the V-net architecture but differs in that it does not have a decoding path; instead fully-connected and softmax layers are added to the end of the encoding path to perform OAR classification. We used 30 images for each structure to train TG263-Net. We then evaluated the accuracy of the model on 2034 correctly labelled structures and 108 incorrectly labelled structures. To ensure robustness of the final label, majority voting was used on predictions from augmentations (n=9) of the test data.
Results: The test accuracy of our model was 99.95%. There was only one out of 2142 cases for which our model did not correctly predict the name of the normal structure. Visual inspection of this case revealed that the original structure was irregularly contoured.
Conclusion: We only needed to curate 30 patientsâ€™ data to correctly label the name of the head-and-neck normal structures for hundreds of retrospective data. This tool will be useful for facilitating big data studies with consistent nomenclature.