Room: Exhibit Hall | Forum 5
Purpose: Accurate segmentation of targets and OARs is the key step for IMRT of nasopharyngeal carcinoma (NPC). This step is extremely time-consuming because the targets and OARs are layered on more than 100 CT slides. In this work, we construct a hybrid convolutional neural network framework to carry out deep learning for segmentation of targets and OARs in NPC CT images.
Methods: FCN8 models on Caffe, an open-source deep learning simulation tool, have been trained by Fine Tune. Based on pre-trained features on the imageNet database and migration techniques, the deep learning process for NPC structures did not require numerous training samples. Fifty NPC patients who received IMRT at the West China Hospital Cancer Center from January 2016 to May 2017 were enrolled to create 3D CT scan datasets with delineations of targets and OARs. Forty of them were used for training and ten of them were used for testing. We found that different anatomic structures need different model parameters to obtain best results. So the classification of slices for location in CT scans for every patient must be executed by Alexnet before segmentation with an appropriate FCN model. The Dice similarity coefficient (DSC) between the contour predicted by trained models and the manually delineation was used to quantitatively analyze the segmentation accuracy of the hybrid FCN8 models.
Results: The targets and OARs that were trained include GTVnx, CTV1, brain stem, spinal cord, temporal lobe, optic nerves, optic chiasm, temporomandibular joints, mandibles, ears and parotid glands. For the 10 test cases, the average DSC of GTVnx was 86.3%, the average DSC of CTV1 was 83.2%, and the average DSC of OARs was 59.7% -88.9%.
Conclusion: Our practice shows that it is feasible to use the hybrid convolutional neural network trained by representative datasets to automatically delineate nasopharyngeal carcinoma targets and OARs.