Z Li¹, Q Zhu¹, Z Li¹, J Fu¹, Z Yang²*, (1) Shanghai Sixth People's Hospital, Shanghai, Shanghai, CN, (2) Duke University, Durham, NC

• 
  Z Yang

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

SU-K-207-4 (Sunday, 7/10/2022) 5:00 PM - 6:00 PM [Eastern Time (GMT-4)]

Room 207

Purpose: Although the dose distribution prediction has been largely applied and researched in External Beam Radiation Therapy, it’s still far relatively unexplored in the realm of brachytherapy. In this study, a dose prediction model for HDR-brachytherapy was proposed using 3D Cascade Squeeze and excite(SE)-Unet.

Methods: A cascade SE-Unet was developed to predict dose distribution given anatomical information, applicator position and dose prescription. Totally, 120 patients with various applicators inserted, including ovoid, tandem, vaginal and needles, were used for training and validation, and another 20 for test.The network input consists of 6 channels, including the patient CT, binary masks for bladder, rectum and HRCTV contours, dose map and applicator mask. The first SE-Unet is used to predict coarse dose distribution and the second for further fine-tuning of the initial results. SE block was used to perform feature recalibration, and thus selectively emphasize informative features and suppress less useful ones. The model performance was evaluated by the residual of clinical-related dosimetric metrics between ground truth and prediction.

Results: The dose deviation between ground truth and prediction for HRCTV, △D90% were -0.04±0.31, for bladder and rectum, △D2cc, were -0.01±0.22, and -0.08±0.26, respectively. The absolute dose error of D90% in HRCTV was 3.83% of the prescription dose, and for D2cc in bladder and rectum were 3% and 3.5%, respectively.

Conclusion: We proposed a cascade SE-Unet model for 3D dose prediction. The proposed framework is capable of generating accurate dose distribution for HDR-brachytherapy.

Keywords

Not Applicable / None Entered.

Taxonomy

TH- Brachytherapy: HDR Brachytherapy

Contact Email