Efficacy of Medical Image Transformations within an Open Sourced Head and Neck Deep Learning Auto-Contouring Framework

Abstract

<h3>Purpose/Objective(s)</h3> Radiation therapy treatment plans have recently utilized deep learning methods to automatically contour organs at risk (OAR). In this study, we use an open-source framework, Medical Open Network for AI (MONAI), and the 3D U-Net network to automatically segment head and neck OAR structures. This study shows that using pre-processing medical image transformations improves performance of auto-contouring. <h3>Materials/Methods</h3> This study used the MONAI deep learning framework to create 28 distinct segmentation models for head and neck OAR structures from a dataset of manually contoured structure sets from CT simulation images. The MONAI framework integrated pre-processing transforms included image windowing, pixel resampling, foreground cropping, and random crop patching that was performed on the training sets prior to model training. Model training was performed using a 3D U-Net algorithm. Mean dice similarity coefficients were computed for each OAR deep learning model. <h3>Results</h3> The training was conducted on a single institution's dataset composed of 609 unique OAR structure sets. Select mean dice similarity coefficients achieved in the auto-contouring model's validation sets ranged as high as 0.906, 0.804, and 0.775 for the orbit, submandibular gland, and thyroid respectively. <h3>Conclusion</h3> We trained a series of deep convolutional neural network models that perform automated contouring of head and neck OAR with a high degree of accuracy, outperforming other models to date. The models were trained using the open-source MONAI framework that is easily accessible and customizable to anyone in radiation oncology. The utility of medically specific image transformations available within MONAI allow for improved model performance and generalizability, and similar techniques could be employed in future studies.