Machine Learning-Derived Input-Function (MLDIF) in Dynamic ¹⁸F-FDG PET

This repository contains source code for non-invasive prediction of the arterial input function (AIF) in small-animal dynamic FDG PET studies.

Two well-known, machine learning-based regression models, have been implemented, based on Gaussian processes (GP) and long short-term memory (LSTM) recurrent neural network, respectively.

Usage

The followings files are available:

• MLDIF lstm_train.py: Train an LSTM model with training and validation data.
• MLDIF_lstm_evaluate.py: Evaluate a trained LSTM model on test data.
• MLDIF_tissue_region_importance_train.py: Train an LSTM model with different feature combinations.
• MLDIF_tissue_region_importance_evaluate.py: Evaluate an LSTM model trained with feature combinations.
• MLDIF_GP.py: Train and evaluate two different GP models

Please see the comments in the code, and the puplication below, for further implementation details.

Citation

Please cite any usage of the content of this repository as:

Kuttner S, Knutsen Wickstrøm K, Kalda G, Dorraji SE, Martin-Armas M, Oteiza A, et al. Machine learning derived input-function in a dynamic 18F-FDG PET study of mice. Biomed Phys Eng Express. 2019 Dec 20;1–46. Available from: https://doi.org/10.1088/2057-1976/ab6496