Deep learning prediction of proton and photon dose distributions for paediatric abdominal tumours

ObjectiveDose prediction using deep learning networks prior to radiotherapy might lead tomore efficient modality selections. The study goal was predict proton and photon dose distributions based on the patient-specific anatomy assess their clinical usage for paediatric abdominal tumours.Material methodsData from 80 patients with neuroblastoma or Wilms’ tumour included. Pencil beam scanning (PBS) (5 mm/ 3%) volumetric-modulated arc therapy (VMAT) plans mm) were robustly optimized internal target volume (ITV). Separate 3-dimensional patch-based U-net trained PBS VMAT distributions. Doses, planning-computed tomography images relevant optimization masks (ITV, vertebra organs-at-risk) of 60 used training a 5-fold cross validation.The networks’ performance evaluated by computing relative error between planned predicted dose-volume histogram (DVH) parameters 20 inference patients. In addition, organs-at-risk mean difference modalities calculated (ΔDmean = DVMAT-DPBS). Two radiation oncologists performed blind PBS/VMAT selection either ΔDmean.ResultsAverage DVH differences ≤ |6%| both modalities. classified Dmean as gain > 0) 98% precision. An identical compared ΔDmean made 18/20 patients.ConclusionDeep accurate tumours established. These allowing fast visualisation aid in identifying optimal technique when experience and/or resources are unavailable.