Artificial Neural Network Algorithm to Cluster and Visualize Phantom Experiment Data

Recent attention has focused on the provision of texture analysis for quantification of intratumor uptake heterogeneity in PET/CT images. This allows the extraction of quantitative features from medical images in the process termed ‘radiomics,’ with the promise of discovering biomarkers that are correlated with end point information (i.e. tumor type, therapy response, prognosis). The conventional complex workflow for calculation of radiomic features introduces numerous confounding variables such as acquisition imaging time (post administration of radiopharmaceutical) and variability in the method of segmentation for region of interest (ROI) identification. Using machine learning techniques for feature (and their combinations) selection can serve as a promising method to alleviate redundancy in radiomics. The application of the self-organizing map to radiomic analysis serves as a powerful general-purpose exploratory instrument to reveal the statistical indicators of texture distributions. For this purpose, texture features from PET-CT images of a within house designed phantom with 4 tumors inserts with different level of heterogeneity were analyzed whilst varying imaging acquisition time and ROI segmentation contour size. The self-organizing map was used to interpret the varying distribution of texture parameters and revealed a distinct cluster of texture features dependent on contour size providing additional evidence that contour size is a confounding variable when performing texture analysis. Future work will explore the incorporation of this revealed dependency in a prediction model in the presence of end point information which may be achieved by utilizing the self-organizing map in clinical data.