X Xie*, H Yan, J Dai, National Cancer Center/National Clinical Research Center for Cancer/Cancer Hospital, Chinese Academy of Medical Sciences and Peking Union Medical College, Beijing, China

• 
  X Xie

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

PO-GePV-M-169 (Sunday, 7/25/2021)   [Eastern Time (GMT-4)]

Purpose: Contouring of breast tumor bed in planning CT is important to post-operative radiotherapy of patients after breast conserving surgery (BCS). However, this contouring is highly uncertain due to specimen volume, seroma size, clarity, surgical clips, inter-observer variability, and other factors. To alleviate this issue, prior information was introduced in a 3D U-Net model to predict the contour of breast tumor bed automatically.

Methods: The pre-operative CT was first aligned to the planning CT via intensity-based B-splines registration. Then the contour of breast tumor expansion in pre-operative CT was transformed to planning CT accordingly. This transformed contour indicates the approximate area of breast tumor bed in planning CT and would be valuable prior information in searching for more accurate contour. Next, the transformed contour with CT images were fed into a 3D U-Net model to predict the fine contour of breast tumor bed. 110 sets of CT images (each including pre-operative CT and planning CT) and the corresponding label images were used for model training with five-fold cross-validation. The dice similarity coefficient (DSC) and Hausdorff distance (HD) were used to evaluate the accuracy of 3D U-Net model.

Results: The average DSC and HD of prediction model were 0.685±0.065 and 20.816±14.565 for breast tumor bed without the input of prior information, while the average values were 0.795±0.038 and 13.896±8.737 with the input of prior information. The differences were statistically significant (0.685 vs. 0.795, P = 0.0014 < 0.05; 20.816 vs. 13.896, P = 0.002 < 0.05).

Conclusion: The introduction of prior information improved the prediction accuracy of breast tumor bed contour. This method provides an automatic way to contour breast tumor bed more accurately in CT image and will be useful in other image modalities.

ePosters

Keywords

Target Localization, Segmentation, Breast

Taxonomy

IM/TH- Image Segmentation Techniques: Modality: CT

Contact Email