Yaoying Liu^1,3,#, Xuying Shang^1,2,3,#, Gaolong Zhang¹, Xiaoyun Le¹, Wei Zhao¹, Zishen Wang², Chunfeng Fang², Baolin Qu³, Shouping Xu^4,* (1) Beihang University, School of Physics, Beijing, China. (2) Hebei Yizhou Tumor Hospital, Department of Radiation Oncology, Hebei, China. (3) The First Medical Center of PLA General Hospital, Department of Radiation Oncology, Beijing, China. (4) National Cancer Center/Cancer Hospital- Chinese Academy of Medical Sciences and Peking Union Medical College, Department of Radiation Oncology, Beijing, China. # Yaoying Liu and Xuying Shang contributed equally to this work.

• 
  
• 
  
• 
  
• 
  
• 
  
• 
  
• 
  

PO-GePV-T-130 (Sunday, 7/10/2022)   [Eastern Time (GMT-4)]

ePoster Forums

Purpose: In proton therapy, an efficient and accurate dose calculation algorithm is a key for qualified plans and delivery. The delivered proton cases were relatively hard to collect. We designed a proton pencil beam algorithm (PBA) boost model: AA-Net trained by a small dataset, effectively improving PBA accuracy.

Methods: Due to patients' variable anatomical structure that can cause PBA dose errors, Anatomical Attention Network (AA-Net) was built based on a novel Anatomical Attention Mechanism, which used CT's anatomical information to focus significant features in the network. Twenty-seven delivered lung cases treated by the proton system have been enrolled in our study: 15 cases used for training and 12 cases used for testing. Each case has paired PBA and Monte Carlo (MC) doses for model training. Conventional U-Net was selected as a comparison model.

Results: MC dose was a reference in dose comparison. For 12 test cases, AA-Net improved gamma passing rates (2mm/2%, 2mm/3% and 3mm/3%, 10% threshold) from PBA's 81.86%, 86.86% and 90.73% to 93.87%, 96.01% and 97.74%, compared with U-Net's 89.05%, 92.98% and 95.52% (p<0.05). Even for 4 cases with large PBA errors, AA-Net also performed effectively, which improved the passing rates (2mm/2%, 10% threshold) from PBA's 60.20% to 90.13%, compared with U-Net's 82.56%. Percentage dose errors of Lung V5, and CTV Dmean, D95, D90, D50, were improved by AA-Net from PBA's 0.17%, 2.53%, 2.19%, 2.04%, 2.67% to 0.14%, 1.3%, 1.39%, 1%, 1.64%, while U-Net performed normally, which even increased the percentage dose errors of PBA dose (p<0.05).

Conclusion: Our anatomical attention-based AA-Net can improve proton PBA dose accuracy, even trained by a small dataset. AA-Net's success was its focus on anatomical information, which can improve the accuracy of the proton PBA. Combining radiotherapy physics and computer science is significant for model designing.

Keywords

Protons, Pencil Beam Algorithms, Monte Carlo

Taxonomy

TH- External Beam- Particle/high LET therapy: Proton therapy – computational dosimetry-deterministic

Contact Email