ESTRO 2022 - Abstract Book

S784

Abstract book

ESTRO 2022

PD-0892 TomoTherapy QA prediction using a combination of complexity metrics and sinograms radiomics features

S. Cavinato 1,2 , A. Bettinelli 1 , F. Dusi 1 , M. Fusella 1 , A. Germani 1 , F. Marturano 1 , M. Paiusco 1 , N. Pivato 1 , A. Roggio 1 , M.A. Rossato 1 , A. Scaggion 1 1 Veneto Institute of Oncology IOV-IRCCS, Medical Physics Department, Padova, Italy; 2 Università degli Studi di Padova, Dipartimento di Fisica e Astronomia ‘G.Galilei’, Padova, Italy Purpose or Objective The purpose of this work was to assess whether a set of newly defined complexity metrics and radiomic features extracted from TomoTherapy plans’ sinograms can enhance the predictive capability of a linear model for the prediction of TomoTherapy ® patient-specific QA results. Materials and Methods A cohort of 636 patients treated between June 2018 and April 2021 using the helical TomoTherapy ® platform (Radixact) was collected, 512 planned using the Precision TPS while the remaining 124 using the RayStation TPS. Three groups of complexity metrics were extracted using the in-house developed Matlab ® (The MathWorks Inc, Natick, MA, USA) library TCoMX [1]: • Group A : 12 typical delivery parameters (e.g. Gantry Period, Treatment Time, etc) • Group B : 14 metrics proposed by Santos and colleagues [2] • Group C : 42 newly developed metrics describing the field geometry and beam modulation [3]. Additionally, 174 radiomics features ( Group D) were extracted from the 2D sinograms using the IBSI compliant software S- IBEX [4]. Separate linear models were realized for the two TPSs to predict the 3%,2mm gamma index passing rate computed with global normalization and a 10% threshold [5]. Four different models were created considering four different sets of variables obtained by progressively adding one group at a time to Group A . A variables pre selection based on the Pearson’s Correlation Coefficient (PCC) was applied to remove strongly correlated variables (|r|>0.8). Each model was trained 1000 times by randomly splitting the dataset into 75% training and 25% validation at each try. In each training phase, variables could be added (removed) progressively to (from) the model through a stepwise approach based on the p-value of an F-statistics. The starting point was always a constant model. The mean absolute error (MAE) and PCC between the true and predicted gamma indexes were computed to measure the performance. Results Figure 1 shows that the progressive inclusion of the new groups of variables increases the average model quality (lower MAE, larger PCC) for both training and validation sets. Only the models obtained for RayStation can be considered partially predictive (PCC > 0.6). Figure 2 shows that groups C and D tend to replace groups A and B suggesting a potentially higher capability of the newer variables to predict the gamma score. In general, the average number of variables included in the model tends to increase as the considered number of groups increases and ranges from 1 (group A only) to approximately 5 (all groups).