Z Zhang*, T Zhao, H Gay, W Zhang, B Sun, Washington University in St. Louis, St. Louis, MO

• 
  Z Zhang

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

TU-F-TRACK 6-7 (Tuesday, 7/27/2021) 4:30 PM - 5:30 PM [Eastern Time (GMT-4)]

Purpose: Manual contouring used in the current treatment planning process is tedious, time-consuming, and can produce inconsistent results. We develop a novel and efficient CNN-based deep learning method for automated multi-organ segmentation on head and neck (HN) CT images.

Methods: The new method uses a densely weaving attention U-net (WAU-net) structure as the backbone and implements a novel boosting-style training strategy. WAU-net follows the basic U-net structure and uses attention fusion blocks for the decoder path. It evolves from shallow structures to more complex ones while forcing itself to exploit deeper features by down-weighting the regions correctly contoured by simpler networks. Under the boosting-style training strategy, the network evolves gradually, stressing previously wrongly labeled pixels to achieve finer segmentation outcomes. We trained our new method to contour 10 OARs within a single forward propagation. We performed 10-fold cross-validation with 90 cases and tested the performance on the rest 25 cases. The accuracy of modeled contours was evaluated with the Dice similarity coefficient (DSC), average Hausdorff distance (AHD), and average surface Hausdorff distance (ASHD) using clinical contours as a reference to the ground truth. The accuracy of the model was compared to other popular deep learning methods.

Results: Our method achieved superior or comparable results to several existing deep learning approaches. In addition, our new model used substantially fewer model parameters and less computation cost. The model also achieved state-of-the-art performance on the testing dataset. The average DSC on the testing data sets were as following: brainstem (0.88), chiasm 0.71), mandible (0.95), spinal cord (0.89), left optic nerve (0.74), right optic nerve (0.75), left sub-mandibular (0.81), right sub-mandibular (0.80), left & right parotid (0.85) and right parotid (0.83).

Conclusion: WAU-net provides state-of-the-arts accurate contouring of multiple OARs, which can potentially facilitate the routine OARs delineation process for HN cancer.

Funding Support, Disclosures, and Conflict of Interest: This work is funded by Varian Medical Systems.

Handouts

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email