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

Area of Science:

• Nuclear medicine
• Medical imaging
• Computational biology

Background:

• Dynamic 18F-fluorodeoxyglucose (FDG) positron emission tomography (PET) coupled with tracer kinetic modeling quantifies glucose metabolism.
• Accurate arterial input function (AIF) determination is crucial for reliable kinetic modeling in FDG PET studies.
• Non-invasive AIF estimation methods are highly desirable, especially in small-animal research where blood sampling is challenging.

Purpose of the Study:

• To evaluate two non-invasive machine learning models, Gaussian processes (GP) and long short-term memory (LSTM) recurrent neural networks, for predicting the AIF in small-animal dynamic FDG PET studies.
• To compare the accuracy of machine learning-derived AIFs against a reference AIF derived from image data.
• To assess the impact of different tissue regions on AIF prediction accuracy.

Main Methods:

• Two machine learning models (GP and LSTM) were trained using 68 dynamic FDG PET/CT mouse scans with 7 delineated tissue regions.
• A reference AIF was generated by fitting a known model to vena cava and left ventricle image data, as blood samples were unavailable.
• Model performance was evaluated by comparing predicted and reference AIFs using area under the curve (AUC) and root mean square error (RMSE). Net influx rate constants (Ki) were calculated and compared using various statistical analyses.

Main Results:

• Both GP and LSTM models generated AIFs with AUCs comparable to the reference AIF.
• The LSTM model exhibited a lower RMSE compared to the GP model, indicating more accurate AIF prediction.
• Calculated net influx rate constants (Ki) using both predicted AIFs showed good agreement with reference values, with no significant differences observed.
• The myocardium was identified as an important region for AIF prediction, though similar RMSE was achieved without it.

Conclusions:

• Machine learning approaches, particularly LSTM, enable accurate and non-invasive prediction of image-derived AIFs in mouse FDG PET studies.
• The LSTM model offers a more accurate alternative to GP for AIF estimation, yielding lower prediction errors.
• These findings support the use of machine learning for robust quantification of glucose metabolism in preclinical PET imaging, reducing the need for invasive blood sampling.