ESTRO 2021 Abstract Book

S1517

ESTRO 2021

Conclusion Although complex models are used to analyze natural image datasets (ImageNet), simpler models can be used to train typical medical image datasets without sacrificing accuracy. In this work, we have validated different techniques to reduce the number of parameters of the models and as such reduce their complexity and training time. The new architectures will be released for their public use. PO-1792 On dose cube pixel spacing pre-processing for features extraction stability in dosomic studies L. Placidi 1 , D. Cusumano 1 , J. Lenkowicz 1 , L. Boldrini 1 , V. Valentini 1 1 Fondazione Policlinico Universitario A. Gemelli IRCCS, Radiation Oncology, Rome, Italy Purpose or Objective Dosomics is a novel texture analysis method to parameterize ROIs and to produce quantitative dose features encoding the spatial and statistical distribution of radiotherapy dose at higher resolution if compared to the standard organ-level dose-volume histograms. The stability of dosomics features extraction on dose cube pixel spacing variation has been investigated in this study. Materials and Methods The 3D dose dataset has been generated considering all the possible combinations of four dose grid resolutions and two calculation algorithms dose calculation of 17 clinical delivered dose distributions (P n ). For each combination, single dose voxel cube has been post-processed considering 4 different dose cube pixel spacing values: 1x1x1 mm 3 , 2x2x2 mm 3 , 3x3x3 mm 3 and the one equal to the planning CT (gs). Dosomics features extraction has been performed with an in-house developed software. Four different ROIs have been considered for the analysis: PTV, the two closest OARs and a RING structure. The stability of each extracted dosomic feature has been analyzed in terms of intraclass correlation coefficient (ICC) and coefficient of variation (CV). Cumulative ICC has been employed to evaluate the variation among the different dose distributions. Results A total of 465664 dosomics features have been extracted from the selected four ROIs, considering the 544 3D dose distributions. For each P n 36992 dosomic features belonging to the STAT family were extracted, 217600 to GLCM family, 141440 to GLRLM family and 69632 to GLSZM family. Stability curves (CV versus the normalized number of features) were analyzed for each P n : in figure 1 is depicted the stability curve of P 1 .