ESTRO 2021 Abstract Book

S655

ESTRO 2021

Conclusion The users appreciated the intuitive process of navigating the trade-off curve to select the best solution for a patient. To improve time efficiency user training is on-going and additional planning criteria need to be implemented to improve plan quality, e.g. dose homogeneity and normal tissue dose constraints. After one year of clinical experience, we can conclude that AI based TP for HDR prostate BT is successfully implemented in our clinic and is a promising tool for further applications in BT. PD-0822 Bespoke vs machine learned: can expert Pareto navigated treatment planning be modelled? I. Foster 1 , P. Wheeler 2 , E. Spezi 1 , J. Staffurth 3 , A. Millin 2 1 Cardiff University, Engineering, Cardiff, United Kingdom; 2 Velindre Cancer Centre, Medical Physics, Cardiff, United Kingdom; 3 Cardiff University, Medicine, Cardiff, United Kingdom Purpose or Objective EdgeVcc is an automated planning solution developed via scripting in RayStation and integrates Pareto navigation based multi-criteria optimisation (MCO) [1]. The MCO interface enables expert observers to intuitively select the optimum treatment plan for a given patient. When optimisation parameters associated with the plan are extracted, a bespoke gold standard plan (GS) for that individual patient can be generated, or parameters utilised within algorithms to inform the planning of novel patients. The purpose of this work was to develop and evaluate a machine learning algorithm (ML), built on the geometric information of anatomy alone, to predict GS optimisation parameters. Materials and Methods 30 prostate cancer patients previously treated at Velindre Cancer Centre were split randomly into 2 cohorts: 20 training and 10 validation. A previously calibrated MCO automated planning protocol was used as a base protocol. Planning parameters included: rectum and bladder D mean , rectum and bladder D max , PTV conformality, and intra-PTV dose falloff. An expert operator used MCO to create patient specific GS plans (MCO GS ). Spatial, volumetric and other derived anatomical features were extracted comprising 27 non- correlated features including distance metrics between regions-of-interest (ROIs), total volume of ROIs and overlap volumes. Linear, quadratic and cubic regression models of up to 5 features (i.e. 15 model architectures) were produced for each parameter and tested using leave-one-out. The optimum architecture was judged to be that which achieved the smallest mean squared error following the leave-one-out analysis. Chosen architecture was then used to generate ML plans (MCO ML ) for 10 validation patients and assessed dosimetrically using Wilcoxon signed rank testing. Results Model architecture chosen for bladder D max was quadratic; all others were linear. All models contained less than three features. The most frequently used features were both spatial: the distance from the centre of PTV48 to the centre of the rectum and the largest distance between PTV48 and the bladder.