Optimising ⁶⁸Ga-SSTR PET radiomic feature extraction towards improved feature stability and predictive classification...

The application of ⁶⁸Ga PET imaging based radiomic analysis to the classification of neuroendocrine tumours (NETs), predicting the response to peptide receptor radionuclide therapy (PRRT) therapy and investigating the heterogeneity of NETs is an area of active research. However, there is limited information available regarding the stability of ⁶⁸Ga PET based radiomic features for clinically relevant volumes and activities or the optimal pre-processing parameters.

Optimisation of ⁶⁸Ga specific radiomic feature extraction and feature selection methods were performed in this study, using a range of uniform and heterogeneous phantoms. Radiomic feature stability was assessed over a range of activities and volumes, for a selection of reconstruction settings and quantisation methods. The ability of an optimised set of radiomic features to generate a prediction model for NET patient tumour grade using machine learning algorithms was then investigated.

The work presented here confirmed that reducing the radiomic feature set prior to clinical model building was beneficial and led to the generation of more accurate clinical prediction models. Quantitative assessment of volume dependency effectively reduced the feature set while maintaining clinically relevant features. Optimisation of the pre-processing quantisation method improved feature stability for small clinically relevant volumes. Five robust ⁶⁸Ga radiomic features were identified that could accurately predict the patient's NET grade when the quantisation parameters were optimised using Rice's rule.

While quantification of small volumes typically seen in NETs remains challenging, optimising the quantisation parameters improves feature stability for volumes > 1.2 cm³.