Room: Exhibit Hall | Forum 2
Purpose: A Nested Unet (Unet++) architecture is proposed for fast and accurate OARs segmentation in Nasopharyngeal Cancer (NPC) for radiotherapy.
Methods: Unet++ is an end-to-end framework, which adopted the added convolution layer and improved skip connection upon the original Unet structure. 147 NPC patients, with supine position and immobilized with thermoplastic mask, have CT scans for treatment planning. The resolution of image reconstruction is 512Ã—512 with 3.0mm thickness. All the OARs listed in NRG_HN001 were delineated by a radiation oncologist and verified independently by another one. 80% patients were randomly chosen as training set for adjusting the parameters, and the remaining 20% cases were regarded as test set for evaluating the proposed method performance. Six organs, including brainstem, spinal cord, left eye and right eye, left parotid, right parotid, were firstly tested on Unet++. The initial learning rate was set to 0.0001, the attenuation factor of learning rate to 0.0008, the attenuation step to 2,500, and the maximum iteration times to 80,000. NVIDIA_GTX2080_GPU was used in the experiment. Dice similarity coefficient (DSC) and Hausdorff distance (HD) were used for Unet++ segmentation evaluation.
Results: It takes about 18 seconds to segment one patient's CT automatically using Unet++, which is greatly improves the efficiency compared with manual delineation. All the OARs DSC values were higher than 0.73 with a mean value of 0.86. The Hausdorff distance values were within 4.1 mm, with a mean value of 3.6 mm, which demonstrated that the proposed method could segment OARs accurately. Eyes have the highest accuracy of automatic segmentation, with a DSC reaching 0.93. The HD value of spinal cord is 3.5mm, with a relative low DSC value.
Conclusion: The Unet++-based method for OARs automatic segmentation achieves an accuracy result, and the Unet++ has the potential to improve the contouring consistency and streamline radiotherapy workflows.
Funding Support, Disclosures, and Conflict of Interest: Student's Platform for Innovation and Entrepreneurship Training Program - 201813902075; 201813902071; 201713902050; Hubei Key Laboratory of Medical Information Analysis and Tumor Diagnosis and Treatment - PJS140011504; Natural Science Foundation of Guangdong Province - 2017A030310217; Pearl River S&T Nova Program of Guangzhou - 201710010162
CT, Segmentation, Radiation Therapy