Room: Karl Dean Ballroom C
Purpose: Due to the complexity of the head and neck organsï¼Œaccurate segmentation of organs-at-risks(OARs) is a crucial and time-consuming process during the planning of radiation therapy for head and neck cancer treatment. In this study, we try to develop a simple deep learning based auto segmentation algorithm to segment OARs on CT images.
Methods: CT scans of 364 patients with head and neck cancer were enrolled in this study. We built the training model with a 2D U-net similar network. Patientsâ€™ image data and manually segmented ROI, including spinal cord, brain stem, left and right temporal lobe, left and right eye, left and right optic nerve, left and right parotid, oral cavity, larynx, and chiasm were inputted into CNN network. We randomly separated dataset into training (90%,) and validation (10%) datasets. Four indices were calculated to evaluate the similarity of automated and manual segmentation, including Hausdorff distance (HD), average surface distance (ASD), Dice index (DSC), and Jaccard index (JSC).
Results: Our method was trained with 334 patientsâ€™ CT images and validated on a set of 30 patientsâ€™ images. The DSC for validation dataset are listed below: spinal cord 0.83Â±0.09, brain stem 0.86Â±0.05, right temporal lobe 0.79Â±0.08, left temporal lobe 0.76Â±0.07, right eye 0.85Â±0.04, left eye 0.85Â±0.05, right optic nerve 0.60Â±0.11, left optic nerve 0.59Â±0.11, right parotid 0.78Â±0.05, left parotid 0.79Â±0.05, oral cavity 0.69Â±0.17, larynx 0.79Â±0.10, chiasm 0.49Â±0.15.It cost about six minutes to finish the segmentation,.
Conclusion: We found that our simple deep learning neural network can quickly and accurately segment OARs after training with a representative database of 400 patientsâ€™ HaN CT images. This shows the great potential of deep learning in radiotherapy treatment planning.