j Peng*, J Wang,

• 
  j Peng

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

PO-GePV-T-263 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: Automated and deep-learning-based planning techniques aim to produce effective and robust radiotherapy plan. Deep-learning-based planning uses a patient-plans library to make a model that can predict achievable 3D dose distribution for new patients and uses dose-volume histograms calculated from the dose distribution to set optimization objectives. We compared a deep-learning-based planning with script-based planning for rectal cancer patients using 3 different goals settings.

Methods: VMAT plans of 51 rectal cancer patients that included sparing of the bladder, left and right femur heads were produced. Deep-learning-based plans and script-based plans were created with 3 different clinical goals setting: easy, moderate and hard. Goals of plan included prescription and max dose of PTV, mean dose of bladder and mean dose of femur head. The PTV prescription was same (50Gy) for 3 goals settings. Hard goals setting required lower PTV-Dmax and more stringent OARs dose objective. Easy goals setting was opposite. Deep-learning-based plans were compared with script-based plans using dose-volume indices: bladder D15/D50, left femur head D25/D40, right femur head D25/D40, and PTV D2/D95. And the clinical acceptable of these plans were accessed with clinical routine criteria.

Results: Compare to script-based plan, deep-learning-based plan for moderate goals setting, decreased bladder D15 by 7.24% and D50 by 33.35%, right femur head D25 by 2.56% and D40 by 6.71%, left femur head by 6.47% and D40 by 11.39% (p<0.05, paired t-test). For hard goals setting, deep-learning-based plan increased the clinical acceptable rate from 78.4% of script-based plan to 84.3%. For easy goals setting, the rate was improved from 47.1% to 82.4%.

Conclusion: Deep-learning-based planning solution were effective and robust. Moderate goals setting results showed that deep-learning-based planning spared most OARs significantly compared with script-based planning. Hard and easy goals setting results showed that deep-learning-based planning could achieve higher clinical acceptable rate than script-based planning.

Keywords

Not Applicable / None Entered.

Taxonomy

Not Applicable / None Entered.

Contact Email