ESTRO 2020 Abstract Book

S188 ESTRO 2020

OC-0351 Deep learning for rectal spacer stratification in prostate boost radiotherapy C. Thomas 1,2 , I. Dregely 2 , I. Oksuz 2 , T. Guerrero- Urbano 2,3 , A. Greener 1 , A. King 2 , S. Barrington 2 1 Guy's and St. Thomas' NHS Foundation Trust, Medical Physics, London, United Kingdom ; 2 King's College London, School of Biomedical Engineering and Imaging Sciences, London, United Kingdom ; 3 Guy's and St. Thomas' NHS Foundation Trust, Clinical Oncology, London, United Kingdom Purpose or Objective For patients undergoing radiotherapy (RT) for prostate cancer, rectal toxicity is a potential serious complication that needs to be considered. Rectal spacing devices have been shown to reduce the incidence and severity of rectal toxicity, however the cost can be prohibitive for healthcare providers. A decision support system to stratify patients based on rectal dose and toxicity risk is therefore desirable. The aim of this retrospective, simulation study was to investigate the use of a deep learning (DL) network to predict rectal dose distributions, associated rectal toxicity risk and rectal spacer requirement in patients who may be suitable for simultaneous integrated boost RT using simulated intra-prostatic lesions. Material and Methods Twenty patients with histologically-proven prostate cancer were enrolled into a Health Research Authority approved clinical study [*****]. For this simulation study, targets and OARs were re-contoured using the methodology from a national dose escalation pilot study. Six simulated lesions were created within the prostate CTV for each patient, to represent dominant intra-prostatic lesions (DILs). As a reference standard, treatment plans were created in Eclipse v13.6 TPS for the 6 DILs for each patient. A Lyman- Kutcher-Burman NTCP model was then used to predict grade 2 (G2) rectal bleeding risk. A hierarchically dense 3D U-net was trained on binary structure maps (PTV60, PTV53, DIL, rectum and bladder) labelled with associated dose distribution, and then was used to predict rectal toxicity risk, using a leave-one-out cross-validation process, and compared to the reference standard. Results Within the cohort of simulated treatment plans, DVH- derived predicted risk of G2 rectal bleeding ranged from 3.7% to 10.4%, with median risk of 7.2%. Taking delineated structures as input, the network predicted rectal toxicity risk with upper and lower 95% confidence limits of 1.3% and -1.6% respectively. Using DL, 82% of patient plans were correctly stratified for rectal spacer insertion based on a proposed protocol of offering the intervention to those with toxicity risk above the population median.

Fig 1. Dose prediction process for representative patient

Fig 2. Bland Altman plot of G2 rectal bleeding prediction errors Conclusion Research into deep learning for radiotherapy dose prediction is expanding, with neural networks proposed as solutions to expedite the RT treatment planning process. Within this work a DL dose prediction network was developed as a decision-support system for patient stratification. This preliminary work demonstrates that such a network could stratify high risk patients for rectal spacer insertion with > 80% accuracy and identify patients close to the stratification threshold. This approach could reduce the need for time and resource intensive treatment planning. Further model development, training and validation is warranted to maximise accuracy prior to clinical implementation.