Deep learning model development for auto-delineation organs at risk on CT simulation images of head and neck cancer

Abstract

Backgrounds: The organs at risk (OARs) delineation process have an important role in treatment planning for head and neck radiation therapy. OARs have been hit by the over limitation of radiation dose. It effects to loses normal function. Usually in the clinic practice OARs are delineated by Radiation Oncologist (RO). Nevertheless, it takes a long time and depending on RO experience.  Lead to variability intra-inter observation.  

Objective: To develop a deep learning model for auto-delineation of OARs on head and neck cancer patients computed tomography images and evaluating the performance of the model. 

Materials and Methods: Head and neck computed tomography simulation images and DICOM structure radiation therapy were used to input data for training deep learning models. The models were developed by using U-Net architecture and Features Pyramid Networks (FPN) architecture. The architectures were modified by using the VGG19 backbone. Fine tune method and Transfer learning method were used to train deep learning models. The models were evaluated by Dice Similarity Coefficient (DSC), 95 percentile Hausdorff Distance (95%HD), The time of model training, and the time of model auto-predictions.  

Results: The averages DSC results of the TUVGG19 model and FFVGG19 were not less than 0.80 all the OARs except DSC of both parotid glands and spinal cord were not less than 0.72. The averages 95%HD were not more than 2 millimeters all the OARs except 95%HD of Mandible and both parotid glands were not more than 4 millimeters. The average time to predict mask was 1.286 and 1.534 seconds per image for the TUVGG19 model and FFVGG19 model, respectively. 

Conclusion: Deep learning models for auto-delineation of OARs developed. It showed good prediction performance. The results were evaluated by DSC value and 95%HD likewise the value of previous studies and fast predictions mask.