Room: Stars at Night Ballroom 2-3
Purpose: Convolutional neural networks (CNN) have shown promise in significantly improving the segmentation accuracy of organs-at-risk (OAR) by training from a dataset with ground-truth contours. However, one issue in the training data is the different sets of contoured organs across different sites, on even on different patients within one site, causing incomplete labeling and posing challenges for training the model. Instead of throwing away data with incomplete labeling, we developed a new training strategy to take the most advantage of all available data.
Methods: 9 OARs were defined as the segmentation goal: brain stem, chiasm, mandible, left and right optic nerves, parotids, and submandibular glands. 48 cases with 3D CT images were used with 32 selected for training and 16 for testing. Out of the 32 training cases, 12 had at least one contours missing. A CNN based on 3D U-Net was used with optimized network structure and loss function. During training, ambiguity was introduced in the loss function to avoid punishing the missing organs. More specifically, with cross-entropy loss function, only voxels from labeled organs were counted; with Dice loss function, the Dice was ignored for missing organs. The model trained using this strategy was compared with the conventional training method using only the data with complete labeling.
Results: The mean Dice scores were: brainstem: 0.87 vs. 0.85, chiasm: 0.58 vs. 0.49, mandible: 0.93 vs. 0.93, left optic nerve: 0.70 vs. 0.65, right optic nerve: 0.72 vs. 0.64, left parotid: 0.85 vs. 0.85, right parotid: 0.83 vs. 0.83, left submandibular gland: 0.81 vs. 0.80, right submandibular gland: 0.78 vs. 0.78, in which the former used all cases with incomplete labeling and shows advantage.
Conclusion: The proposed training strategy can overcome the obstacles of incompleteness in the training cases and take the most advantage of all available data.