Machine learning derived input-function in a dynamic ¹⁸F-FDG PET study of mice

Abstract

Tracer kinetic modelling, based on dynamic 18 F-fluorodeoxyglucose (FDG) positron emission tomography (PET) is used to quantify glucose metabolism in humans and animals. Knowledge of the arterial input-function (AIF) is required for such measurements. Our aim was to explore two noninvasive machine learning-based models, for AIF prediction in a small-animal dynamic FDG PET study.7 tissue regions were delineated in images from 68 FDG PET/computed tomography mouse scans. Two machine learning-based models were trained for AIF prediction, based on Gaussian processes (GP) and a long short-term memory (LSTM) recurrent neural network, respectively. Because blood data were unavailable, a reference AIF was formed by fitting an established AIF model to vena cava and left ventricle image data. The predicted and reference AIFs were compared by the area under curve (AUC) and root mean square error (RMSE). Net-influx rate constants, K i , were calculated with a two-tissue compartment model, using both predicted and reference AIFs for three tissue regions in each mouse scan, and compared by means of error, ratio, correlation coe cient, P value and Bland-Altman analysis. The impact of di↵erent tissue regions on AIF prediction was evaluated by training a GP and an LSTM model on subsets of tissue regions, and calculating the RMSE between the reference and the predicted AIF curve.Both models generated AIFs with AUCs similar to reference. The LSTM models resulted in lower AIF RMSE, compared to GP. K i from both models agreed well with reference values, with no significant di↵erences. Myocardium was highlighted as important for AIF prediction, but AIFs with similar RMSE were obtained also without myocardium in the input data.Machine learning can be used for accurate and non-invasive prediction of an image-derived reference AIF in FDG studies of mice. We recommend the LSTM approach, as this model predicts AIFs with lower errors, compared to GP.

Cerebral blood flow measurements with 15O-water PET using a non-invasive machine-learning-derived arterial input function

“…In this study, different MLIF models were evaluated with 15 O-labeled water. In our previous research 29 a machine learning approach was also feasible for AIF prediction using FDG, although not yet evaluated in clinical data. Thus, we suggest that the method can be adopted to other tracers by merely training similar MLIF models.…”

“…28 In our previous work, we developed and validated a machine-learning-based input function for 18 Ffluorodeoxyglucose (FDG) in a mouse PET cohort. 29 In short, two learning models were evaluated that predicted an AIF from time-activity curves of up to 7 different tissue regions as input. The main limitation with our previous study was the lack of an AIF, thus the models could only be validated against a reference IDIF.…”

Towards quantitative small-animal imaging on hybrid PET/CT and PET/MRI systems

“…Noteworthy, other non-invasive approaches such as reference tissue models, population-based input functions as well as accumulated activity in the bladder or liver have been optimized to derive the IF for small rodents [183][184][185][186]. More recently, Kuttner et al explored two machine learning methods based on Gaussian processes and long short-term memory to estimate the IF from PET/CT images of mice [187]. The IF generated by both models showed good agreement with the image-derived reference arterial IF generated through fitting a well-established model to vena cava and left ventricle of mice PET scan [187].…”

“…More recently, Kuttner et al explored two machine learning methods based on Gaussian processes and long short-term memory to estimate the IF from PET/CT images of mice [187]. The IF generated by both models showed good agreement with the image-derived reference arterial IF generated through fitting a well-established model to vena cava and left ventricle of mice PET scan [187].…”

A deep neural network for parametric image reconstruction on a large axial field-of-view PET

“…Moreover, parametric image reconstruction methods require an accurate estimation of arterial input function (AIF), for which an invasive blood sampling through a catheter in arterial or arterialized venous [17] was performed in early research, but it is invasive and costly for patient and clinical staff. Therefore, several alternative non-invasive methods have been proposed, including the population-based [18], factor analysis [19], image-driven input function (IDIF) [20][21][22], simultaneous estimation [23] and recent machine learning methods [24]. In IDIF, the most common noninvasive method, the activity distribution of like ascending or descending aorta, and left ventricle (LV) should be aware.…”