Y Cao*, A Vassantachart, O Ragab, J Cui, A Zhuang, Z Shen, S Balik, M Gribble, E Chang, W Yang, University of Southern California, Los Angeles, CA

• 
  Y Cao

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

TU-E-TRACK 6-1 (Tuesday, 7/27/2021) 3:30 PM - 4:30 PM [Eastern Time (GMT-4)]

Purpose: Pre-implant MRI is the gold standard on image-guided brachytherapy for cervical cancer. However, HDR planning is typically done on post-implant CT-based high-risk CTV (HR-CTV) because transferring of pre-implant MR-based GTV to post-implant planning CT is difficult due to large anatomical changes. The goal of this study is to train a dual-path convolutional neural network (CNN) for automatic segmentation of HR-CTV on post-implant planning CT.

Methods: Pre-implant T2-weighted MR images and post-implant CT for 55 (48 for training and 7 for testing) patients were retrospectively solicitated.  MR was aligned to the corresponding CT using rigid registration. HR-CTV and GTV were manually contoured on CT and MR by an experienced radiation oncologist. All images were then resampled to a spatial resolution of 0.5x0.5x1.25 mm. A 3D CNN architecture was constructed with two encoding paths based on CT (no mask) and MR (masked by GTV). Each path used the same 3x3x3 kernel to capture the features from CT and MR. As a comparison, single path CNN with only CT input was trained. Voxel-based dice similarity coefficient (DSC), sensitivity, precision, and Hausdorff distance (HD) were used to evaluate model performance. Cross-validation was performed to assess model stability. 

Results: The dual-path CNN model achieved average DSC of 0.76 (0.70-0.81), Sensitivity of 0.82 (0.79-0.86), and Precision of 0.85 (0.83-0.90). The single-path CNN with only CT input achieved DSC of 0.65 (0.55-0.76), Sensitivity of 0.76 (0.68-0.80), and Precision of 0.78 (0.75-0.83). The area under the curve (AUC) values were 0.83 (0.82-0.84) and 0.76 (0.75-0.77) for dual-path and single-path models, respectively. The HD for the two models were 7.6 mm (4.2-10.5 mm) and 8.3 mm (5.1-12.5 mm).

Conclusion: A CNN model with two encoding paths from pre-implant MR and post-implant CT was successfully developed for automatic segmentation of HR-CTVs for tandem-and-ovoid brachytherapy patients.  

Handouts

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email