Deep Learning Dose Prediction as a Tool to Evaluate Treatment Complications Based On NTCP Models
C Draguet1,2*, A Barragan Montero1, P Populaire2,3, G Defraene2, M Thomas2,3, K Haustermans2,3, J Lee1, E Sterpin1,2, (1) UCLouvain - Institut de Recherche Experimentale et Clinique - Molecular Imaging Radiotherapy and Oncology (MIRO), Brussels, BE, (2) KULeuven - Department of Oncology - Laboratory of Experimental Radiotherapy, Leuven, BE, (3) University Hospital Leuven, Department of Radiation Oncology, Leuven, BE
Presentations
SU-F-206-5 (Sunday, 7/10/2022) 2:00 PM - 3:00 PM [Eastern Time (GMT-4)]
Room 206
Purpose: Dose prediction with deep learning (DL) models is becoming prominent for autonomous radiation treatment planning. Combined with normal tissue complication probability (NTCP) models, dose prediction with DL has the potential to automatically identify patients who are at high risk of toxicity. The goal of this study is to assess the accuracy of our DL dose prediction models for intensity-modulated radiation therapy (IMRT) and pencil beam scanning (PBS) treatments when chained with NTCP models to identify esophageal cancer patients who are at high risk of developing complications and who should be redirected to proton therapy (PT).
Methods: The DL model is based on the U-Net architecture with dense connections. Two models were created: one for photon (XT) plans and one for proton (PT) plans. They were trained on a database of 40 patients using cross validation and a circulating test set to predict the dose distribution for each patient. These models were combined with a NTCP model for postoperative pulmonary complications. The NTCP model used the mean lung dose, age, histology type, and body mass index as predicting variables. The accuracy of our DL models for patient referral is then assessed using ΔNTCP thresholds between XT and PT plans. According to the Dutch Society for Radiotherapy and Oncology, a ΔNTCP ≥ 10 % for grade 2 complications is required for redirecting patients to PT.
Results: Our models succeed in predicting dose distributions, with a mean error on the MLD equals to 0,562 Gy for XT and to 0,374 Gy for PT. For patient referral, we obtained 100% accuracy. The mean residuals (ΔNTCP ground truth - ΔNTCP predicted) is equal to 1,408 %.
Conclusion: This study evaluates our DL dose prediction models in a broader patient referral context and demonstrates their ability to support clinical decisions.
Keywords
Taxonomy
IM/TH- Informatics: Informatics in Therapy (general)
Contact Email