Page 212 - Quantitative Imaging of Small Tumours with Positron Emission Tomography

bed position. We observed that PVC does have a rather small negative impact on the test-retest repeatability, but that this worsened when reducing the counts in the PET image. Therefore, shortening the acquisition time propagates the negative impact of PVC on PET precision. Extending acquisition times currently used for [18F]DCFPyL and [18F]FDHT will not likely be clinically feasible. Still, it should be maximized as the repeatability of uptake quantification can be substantially affected when count density is reduced, as we show in Chapter 4 for [18F]FDHT (25). Taken together, for [18F]DCFPyL and [18F]FDHT PET quantification, the original (non-PVC, EARL1-calibrated) images may be preferred for response assessment studies. Clinical protocols should preferably include double image reconstruction: EARL1 for quantification in response assessment, EARL2 for routine visual analysis (26). Of note, we did observe that the repeatability of whole body tumour burden assessments, which are usable in 177Lu-PSMA response monitoring, are robust to PVC and require no additional image reconstructions. The relative changes measured during treatment might not necessarily lead to different conclusions between non-PVC and PVC images, as we observed in Chapter 3 for lung cancer and Chapter 9 for prostate cancer (7,27). The negative impact of PVC on response assessment will of course depend on the treatment- induced effect sizes. Extracting PET radiomics from small tumours is feasible and valuable In recent years, artificial intelligence (AI) has gained immense popularity and attention in the field of radiology and nuclear medicine (28). While AI models and algorithms have existed for some decades, computational power and large amounts of available data now allow for its potential to be fully revealed (28). Specifically in the field of medical imaging, there are high hopes for AI to of significant clinical benefit, as large amounts of data are acquired on a routine basis. Here, AI may be used to perform physician tasks (e.g. detection or classification of disease), or to make objective predictions on patient outcomes (clinical or pathological). The latter may be of particular interest for PET imaging, as it is inherently quantitative and targeted at tumour biology. In Chapter 6 we investigated the use of artificial intelligence (AI) in analysis of [18F]DCFPyL PET images to predict prostate cancer risk, combining radiomics with machine learning. From [18F]FDG PET studies, it is well known that PET radiomics features, especially those based on texture analysis, are strongly 11 Summarizing discussion 211