Organ At Risk Segmentation In Head And Neck Based On Deep Learning And Anatomical Prior Knowledge

Head and neck cancer is one of the most common cancers worldwide.Image guided radiation therapy(IGRT)is the most effective treatment for head and neck cancer.The successful implementation of IGRT requires accurate delineation of organ-at-risk(OAR)on computed tomography(CT)images.OARs are manually segmented by the oncologist,which is time consuming,laborious,and subjective.Deep learning method is a popular machine learning method in recent years,which automatically extracts meaningful features.Many studies have shown that deep learning was successful and prospective in head and neck OAR segmentation.However,most studies ignored the important role of anatomical prior knowledge in guiding OAR segmentation.The parameters of the designed networks in these studies were huge,costing a lot of computing resources.It is difficult to deploy these methods in hospital.Therefore,we designed a deep learning lightweight model and used anatomical prior knowledge for fast and accurate segmentation of OAR in head and neck cancer.We proposed a three-dimensional(3D)cascade segmentation network with less parameters for segmentation.Firstly,a head and neck OAR template was selected and the region of interest(ROI)of different OAR was designed.We designed a 3D registration network to align the fixed OAR template to the new image for locating OAR in new image.The center distance loss was implemented for improving the locating accuracy.We designed a ROI selection layer,which crops different ROIs based on registration results and OAR volume information.Different ROIs were fed into a 3D multi-view segmentation network for OARs segmentation.The 3D multi-view segmentation network combined the 3D information and two-dimensional information of the CT image.ROI classification branch was added to improve the segmentation accuracy.In addition,we used contextual information to further improve the accuracy of OAR segmentation and reduce OAR positioning error.Finally,we used the shape information to modify the segmentation results.The median filter and hole filling were used to further improve the accuracy of segmentation results.We evaluated the segmentation performance of the proposed method based on two datasets.The average Dice similarity coefficient on the MICCAI 2015 dataset is 79.1%,and the average 95% Hausdorff distance is 3.5 mm.The time cost of segmenting the 9 OARs of a patient with the proposed method was approximately 3.5 seconds.The total parameters of our proposed network were about 1.6 million.It takes 3 minutes for an oncologist with moderate experience to delineate a patient with 9 OARs based on our segmentation results.Compared with the 20～30 minutes required for direct delineation,the delineation time is greatly reduced.We successfully designed the lightweight model for fast and accurate automatic segmentation of OAR in head and neck cancer.The proposed method is expected to be able to actually assist clinicians in OAR delineation in the future,to alleviate clinicians’ manual efforts,and to improve the efficiency and accuracy of OAR delineation.