Purpose Our aim was to develop and validate a machine learning (ML)-based approach for interpretation of I-123 FP-CIT SPECT scans to discriminate Parkinson's disease (PD) from non-PD and to determine its generalizability and clinical value in two centers. Methods We retrospectively included 210 consecutive patients who underwent I-123 FP-CIT SPECT imaging and had a clinically confirmed diagnosis. Linear support vector machine (SVM) was used to build a classification model to discriminate PD from non-PD based on I-123-FP-CIT striatal uptake ratios, age and gender of 90 patients. The model was validated on unseen data from the same center where the model was developed (n = 40) and consecutively on data from a different center (n = 80). Prediction performance was assessed and compared to the scan interpretation by expert physicians. ResultsTesting the derived SVM model on the unseen dataset (n = 40) from the same center resulted in an accuracy of 95.0%, sensitivity of 96.0% and specificity of 93.3%. This was identical to the classification accuracy of nuclear medicine physicians. The model was generalizable towards the other center as prediction performance did not differ thereby obtaining an accuracy of 82.5%, sensitivity of 88.5% and specificity of 71.4% (p = NS). This was comparable to that of nuclear medicine physicians (p = NS). Conclusion ML-based interpretation of I-123-FP-CIT scans results in accurate discrimination of PD from non-PD similar to visual assessment in both centers. The derived SVM model is therefore generalizable towards centers using comparable acquisition and image processing methods and implementation as diagnostic aid in clinical practice is encouraged.

No abstract

Funding Acknowledgements Type of funding sources: None. Background and purpose Myocardial blood flow (MBF) measurements using PET are increasingly used to guide the management of patients with (suspected) coronary artery disease (CAD). Day-to-day variability of these measurements is poor with a 21% standard deviation or 40% 95%-confidence interval [Reference: JACC Cardiovasc Imaging, 2017;10(5):565]. This limits clinical applicability in diagnosis, risk stratification and follow-up as these all depend on comparison of flow values with fixed cut-off values. We expect that reproducibility can be improved by combining flow measurements with the variation of flow values within the myocardium. As entropy is a measure of variability of the associated distribution, we compared the reproducibility of an entropy-based flow parameter with that of conventional myocardial flow reserve (MFR) measurements. Methods We performed a study using intra-individual comparison in 24 patients who underwent rest and regadenoson-induced stress myocardial perfusion imaging using Rubidium-82 on two different PET systems (PET1: Discovery 690, GE Healthcare, and PET2: Vereos, Philips Healthcare) within 3 weeks. MBF for both rest and stress was calculated using Lortie’s one-tissue compartment model (Corridor4DM, INVIA). MFR (ratio of MBF stress/rest) was determined for the myocardial as a whole (MFRglobal), for the three vascular territories: LAD, LCX and RCA (MFRregional) and for the 17 segments. Next, we calculated Shannon’s entropy to measure the variation of the 17 MFR segmental values. We multiplied Shannon’s entropy by the mean of the MFR segmental values resulting in an entropy-based MFR (MFRentropy). For each patient MFRglobal, MFRregional and MRFentropy were compared between both PET systems. For each of the three parameters the test-retest precision was calculated as the SD of the relative difference between measurements. Results The mean difference in MFR measurements between both cameras did not differ from zero (p > 0.05). Mean values for PET1 were MFRglobal = 2.4, MFRregional = 2.4 (LAD), 2.4 (LCX) and 2.5 (RCA), and MFRentropy = 2.4. For PET2 we found MFRglobal = 2.5, MFRregional = 2.5 (LAD), 2.4 (LCX) and 2.6 (RCA), and MFRentropy = 2.5. Test-retest precision was lower for MFRentropy with 11% compared to that of MFRglobal (21%), MFRregional LAD (22%), MFRregional LCX (23%) and MFRregional RCA (24%) (p < 0.01). Conclusion The reproducibility of myocardial flow reserve measurements using Rubidium-82 PET improved by a factor of 2 when an entropy-based flow parameter instead of global or regional MFR parameters is used. This entropy-based flow-parameter may be used to better discriminate ischemia from non-ischemia and may therefore improve CAD management.

Funding Acknowledgements Type of funding sources: None. Background Accurate risk stratification in patients with suspected stable coronary artery disease (CAD) is essential for choosing an appropriate treatment strategy but remains challenging in clinical practice. Purpose Our aim was to develop and validate a risk model to predict the presence of obstructive CAD after Rubidium-82 PET and a coronary artery calcium score (CACS) scan using a machine learning (ML) algorithm. Methods We retrospectively included 1007 patients without prior cardiovascular history and  a low-intermediate pre-test likelihood, referred for rest and regadenoson-induced stress Rubidium-82 PET combined with a CACS scan. Multiple features were included in the ML model; PET derived features such as summed difference score and flow values, CACS, cardiovascular risk factors (cigarette smoking, hypertension, hypercholesterolemia, diabetes, positive family history of CAD), medication; age; gender; body mass index; creatinine serum values; and visual PET interpretation. An XGBoost ML algorithm was developed using a subset of 805 patients to predict obstructive CAD by using 5-fold cross validation in combination with a grid search. Obstructive CAD during follow-up was defined as a significant stenosis during invasive coronary angiography, a percutaneous coronary intervention or a coronary artery bypass graft procedure. The ML algorithm was validated with unseen data of the remaining 202 patients. Results Application of the XGBoost algorithm resulted in an area under the curve (AUC) of 0.93 using the training data (n = 805) and an AUC of 0.89 using the unseen data (n = 202) in predicting obstructive CAD. The strongest predictors were the CAC-scores and quantitative PET derived features. The classical risk factors and medication hardly provided an added value in the prediction of obstructive CAD. Conclusion The developed ML algorithm is able to provide individualized risk stratification by predicting the probability of obstructive CAD.  Although validation with a larger dataset could result in a more well defined performance range, this model already shows potential to be implemented in the diagnostic workflow.