No abstract

The purpose of this work was to assess the capability of radiomic features in distinguishing PET image regions with different uptake patterns. Furthermore, we assessed the stability of PET radiomic features with varying image reconstruction settings. An in-house phantom was designed and constructed, consisting of homogenous and heterogenous artificial phantom inserts. Four artificially constructed inserts were placed into a water filled phantom and filled with varying levels of radioactivity to simulate homogeneous and heterogeneous uptake patterns. The phantom was imaged for 80 min. PET images were reconstructed whilst varying reconstruction parameters. The parameters adjusted included, number of ordered subsets, number of iterations, use of time-of-flight and filter cut off. Regions of interest (ROI) were established by segmentation of the phantom inserts from the reconstructed images. In total seventy eight 3D radiomic features for each ROI with unique reconstructed parameters were extracted. The Friedman test was used to determine the statistical power of each radiomic feature in differentiating phantom inserts with different hetero/homogeneous configurations. The Coefficient of Variation (COV) of each feature, with respect to the reconstruction setting was used to determine feature stability. Forty three out of seventy eight radiomic features were found to be stable (COV ≤ 5%) against all reconstruction settings. To provide any utility, stable features are required to differentiate between regions with different hetro/homogeneity. Of the forty three stable features, fifteen (35%) features showed a statistically significant difference between the artificially constructed inserts. Such features included GLCM (Difference average, Difference entropy, Dissimilarity and Inverse difference), GLRL (Long run emphasis, Grey level non uniformity and Run percentage) and NGTDM (Complexity and Strength). The finding of this work suggests that radiomic features are capable of distinguishing between radioactive distribution patterns that demonstrate different levels of heterogeneity. Therefore, radiomic features could serve as an adjuvant diagnostic tool along with traditional imaging. However, the choice of the radiomic features needs to account for variability introduced when different reconstruction settings are used. Standardization of PET image reconstruction settings across sites performing radiomic analysis in multi-centre trials should be considered.

Medical imaging plays an essential role in the diagnosis and treatment of many types of cancer. Currently, medical images are assessed visually by radiologists and clinicians. However, the full utility of information contained within medical images has yet to be fully explored. One avenue for this exploration is the utilization of "radiomic features" through the application of texture analysis. The numerous radiomic features proposed may vary with confounding variables such as the time post injection of image acquisition and the accuracy of the delineation of the prescribed segmentation volume. To this avail, we propose using the determinant of the correlation matrix to analyze radiomic features robustness to confounding variables. For this purpose, dynamic pre-clinical positron emission tomography (PET) images of 8 mice with mammary carcinoma xenografts (4T1) were binned into 5 minutes intervals from 50 to 70 minutes post injection. The effect of variation in segmentation was also explored by incrementally increasing segmentation volume. From each image set, we extracted 78 Radiomic features for analysis. Analysis. The statistical association measured by the determinant of the correlation matrix when considering contour size was 0.02378; for acquisition time this value was 0.13296. From this analysis we conclude that both temporal variation and segmentation effect the measurement of temporal features and that texture features are less robust to varying acquisition time than to varying segmentation volume.

Texture analysis for quantification of intratumor uptake heterogeneity in PET/CT images has received increasing attention. This allows the extraction of a large number of 'radiomic' features to be correlated with end point information such as tumor type, therapy response, prognosis. The conventional complex workflow for calculation of texture features introduces numerous confounding variables. This non exhaustively includes, imaging time post administration of radiopharmaceutical and the method and extent of functional volume segmentation. A lack of understanding on the dependency of texture features with these variables serves as a detriment to the urgent need to standardize texture measurements to pool results from different imaging centers. The utilization of machine learning techniques for feature (and their combinations) selection serves as a promising method to alleviate redundancy in radiomics. To this avail, we introduce for the first time the application of a Kohonen self-organizing feature map to identify the emergent properties present when performing texture analysis. 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 8 pre-clinical mice with mammary carcinoma xenografts were analyzed with varying post injection imaging time and tumor segmentation contour size. This varying distribution of texture parameters were interpreted by the selforganizing map to reveal two distinct clusters of texture features which are dependent on contour size, providing additional evidence that contour size and hence segmentation method is a confounding variable when performing texture analysis. Furthermore, the self-organizing map can be utilized as a method to incorporate this revealed dependency in a prediction model in the presence of end point information, which will be an area of future work.