Room: AAPM ePoster Library
Purpose: Exhaustive annotation is crucial for deep learning based image segmentation. However, medical imaging data with noisy annotation is ubiquitous and its effects on deep-leaning based image segmentation is unknown. We will study the effects of noisy annotation in the context of deep-learning-based mandible segmentation from CT images.
Methods: CT images of 202 Head and Neck cancer patients are collected. Mandible, is roughly annotated (noisy label) by one of 12 dosimetrists as radiotherapy avoiding structure and further corrected by a physician as clean label. The HighRes3D net is chosen as the segmentation network with 180 images for training, 10 for validation and 12 for test. The ratio of noisy labels (0%, 10%, 20%, 30%, 40%, 50%, 60% and 100%) in the training set is used as independent variable for the robustness studies. The results are assessed from Dice coefficient (DICE) and voxel-wise false positive rate (vs-FPR).
Results: In general, a deep network trained with noisy labels performs inferior than that trained with clean labels. However, no significant difference (two sample t-test, p= 0.25) of metrics is found between the model trained with clean labels (DICE, 0.89±0.07; vs-FPR, 0.08±0.11) and the model trained with 10% noisy labels (DICE, 0.89±0.07; vs-FPR, 0.08±0.10) or with 20% noisy labels (DICE, 0.87±0.08; vs-FPR, 0.08±0.10).
Conclusion: This study suggests that deep learning based medical image segmentation is robust to some extent with regard to noisy annotation in radiation oncology. It may imply that the quantities of labels may mitigate the effects of label inaccuracy for deep-learning based application.
Funding Support, Disclosures, and Conflict of Interest: This work is supported by NIH R01 CA235723 and the seed grant of Radiation Oncology Department at University of Texas Southwestern Medical Center.