Page 140 - Quantitative Imaging of Small Tumours with Positron Emission Tomography

may be of particular benefit for low-grade prostate cancer lesions that tend to be less avid on PSMA PET. We observed that PVC had a substantial impact on most radiomics features (Fig. 6.2A). Also, use of PVC tended to increase the predictive value of radiomics for LNI and any metastasis by improving model stability. Between prediction of metastatic disease and histopathological features there was no single delineation threshold optimal for both, which may be due to the different importance of features between these outcomes. For metastasis prediction, a higher delineation threshold seemed to benefit model performance, while for GS and ECE prediction no clear trend was observed between the different thresholds and use of PVC. Specifically for multicenter settings, harmonization of tumor delineation method and use of PVC data will be crucial. In order to facilitate radiomics analysis, it may be an option to extract radiomics features using a 70%peak threshold on PVC-images for all predictions as this approach tended to improve LNI and metastasis prediction AUCs and model stability (Fig. 6.3), and had minimal effect on the other outcome predictions. Some studies have observed that in radiomics analyses, calculation of 6 textural features might be biased in small tumors or provide little added value above lesion volume itself (41,42), suggesting small lesions might need to be excluded from such studies. Still, the redundancy of those features will depend on a complex relationship between lesion size distributions, level of correlation between the individual features, and the relative importance of those features within the prediction models. Perhaps, a better approach to determine the clinical added value of small tumor PET radiomics might be to determine its predictive value and benchmark this against that of basic PET features. Also, a potential benefit of PVC needs to be considered. Despite analyzing predominantly small lesions (see Supplemental Fig.), we did find significant predictive value in the radiomics data, with (non-significantly) higher AUCs than based on standard PET-metrics. Also, use of PVC seemed to benefit LNI and metastasis prediction. Hence, small tumor radiomics and use of PVC may indeed allow for worthwhile radiomics studies in cancers with predominantly small lesions. Still, future multicenter external validation is needed to demonstrate true benefits of PSMA- radiomics over standard PET metrics in these small prostate cancer lesions, especially since using different PET systems with potentially different imaging protocols might negatively affect radiomics-based predictions more than those based on standard PET features. PSMA radiomics and machine learning 139