Session Item

The purpose of this study is to apply a retrospective exploratory MR-based radiomics and dosiomic analysis based on machine-learning technologies and statistical analysis, to investigate imaging-based biomarkers of clinical outcomes in paediatric patients affected by medulloblastoma. This work was developed in the framework of the INFN-funded Artificial Intelligence in Medicine project.

A database of 50 paediatric patients who underwent surgery, chemotherapy and radiotherapy (RT) was retrospectively selected. It includes information on histology, prescribed drugs and planned dose distributions. Moreover, all MR and CT images acquired from pre-treatment to the end of follow-up are available.
As a first step in a wider and deeper research, multiparametric data were considered, including: last MRI just before RT (T1w, T2w and FLAIR image sets) and dose distribution of the radiotherapy plan. Second order features from those images were extracted with PyRadiomics and analysed with two different programs, a homemade script and RadAR (Radiomics Analysis with R).
Principal Component Analysis technique was exploited to decrease the number of variables involved while maintaining 90% of the variability of the data. Ten features were associated with radio-induced toxicity occurrence. Random Forest classifier was trained on different combinations of the available data. RadAR performs analysis of radiomic features, implementing multiple statistical methods. To compare multiple features together, scaling and normalization were applied. An unsupervised analysis based on hierarchical clustering was used together with the Fisher’s exact test to estimate statistical significance.

We trained and evaluated the machine learning performance within a 5-fold cross validation scheme. At each iteration, a different test set consisting of an equivalent number of subjects with and without toxicity was used. Individual dose, T1w, T2w and FLAIR data were not predictive, leading to average accuracies of 0.44±0.21, 0.56±0.16, 0.59±0.16 and 0.41±0.07, respectively. While combining all multiparametric information a better prediction performance was obtained: 0.62±0.09.
Additionally, results were obtained with statistical features analysis. In fig.1 is reported the correlation matrix used to evaluate the degree of redundancy of feature data in the dataset. The result of Fisher’s test is displayed by Kaplan-Meier plot in fig.2. All the features available were predictive with a specificity of 0.79 and sensitivity of 0.65.

Promising results have been obtained applying Random Forest classifier and statistical analysis to radiomic and dosiomic features of paediatric patients affected by. This multimodal and multiparametric approach could have a large impact for precision medicine, as radiomic biomarkers are non-invasive and can be applied to imaging data that are already acquired in clinical settings.