Room: Exhibit Hall | Forum 5
Purpose: Accurate segmentation of organs at risk (OARs) in head and neck CT images is a fundamental task for head-and-neck cancer radiotherapy. However, manual delineation of OARs is tedious, time consuming and inconsistent. On the other hand, the interpatient OARs variation make accurate automated segmentation a more challenging task. To overcome the challenge, a novel 3D U-Net based network is proposed for automatic head and neck OARs segmentation.
Methods: U-Net is a convolutional network developed for medical segmentation. For the head-and-neck segmentation problem, 3D image segmentation was performed using 3D convolutional layers with a weighted dice loss function to address the class imbalances. To prevent overfitting, a large variety of data augmentation techniques were utilized. Then, graphical model was employed as a label regularizer to further refine the segmentation results produced by the network.
Results: For training the proposed network, 40 head and neck CT scans provided by Public Domain Database for Computational Anatomy (PDDCA) were utilized. Then, the proposed algorithm was validation on the segmentation of brainstem, mandible, left and right parotid with 10 CT scans. And dice coefficient(DC), positive predictive value(PPV), sensitivity, and average surface distance(ASD) were calculated to quantitatively evaluate the performance of the proposed method. An average DC=0.86(brainstem), DC=0.90(mandible), DC=0.78(left parotid), and DC=0.79(right parotid) were obtained.
Conclusion: Experiments on clinical datasets of head and neck patients show that the proposed algorithm can provide useful segmentation for several head-and-neck OARs based on the CT images in minutes. The proposed method thus may be used to improve radiation therapy treatment planning efficiency and consistency.