GAMMA RADIOMICS APPLIED TO RADIOTHERAPY

Introduction: In Radiotherapy the standard method of comparing dose distribution delivered by the linear acelerator to that calculated by the treatment planning system is the gamma function. This method takes only geometrical factors involving dose and distance deviations - in consideration and, despite being largely used, it is a questionable tool. In this context, it would be useful to have an alternative methodology to assure its Results: Scalco, E, in his research, analyzed quality assurance (QA) of Intensity Modulated Radiation Therapy (IMRT) patients, field by field, inserting random errors and extracted radiomics from gamma distribution images to predict if these errors would be detected with the new methodology. As an extension to Scalco, Es work, we analyzed QA from IMRT patients, using radiomics in gamma distribution images, considering all fields at once, to predict QA approval. Goal: Finding a predictive model using artificial intelligence (AI) tools to validate QA acceptance, from images feature analyzed in gamma distribution maps and extracted by radiomics. Methods: For gamma analysis we used criteria 2% / 2mm and minimal approval of 95%. 156 gamma distribution images using the OminiPro I’mRT software were selected. To extract texture features the software PyRadiomics was used. The Gaussian behavior of each Radiomic were analyzed and a T Student test was made to assess whether there was a statistically significance. We inserted the characteristics that presented statistical significance in a random tree. We use k-fold cross validation to train the tree. Results: The characteristics were analyzed by the normality test when combined the same feature for two classes. If they showed a gaussian curve then we performed a T-Student test. Six characteristics showed statistically significance and were used in the model. We analyzed the dispersion in pairs to conclude that they are nonlinearly separable, so we used a random forest to predict the result of the gamma analysis. The data were acquired using k fold cross validation. After 10 folds, the mean accuracy is AUC: 0.837 ± 0.055. Conclusions: A predictive model using random forest algorithm was developed to validate QA acceptance with AUC 0.837± 0.055 using radiomics. 101 images features were extracted but only six of them showed statistically significance and were used in the model. This way it was possible to achieve the same results in the gamma QA approvance using a different analysis.

 

MissingLinkText

No conflict of interest has been declared by the author(s).

Publication History

Article published online:
23 October 2019