TL;DR
This paper introduces a novel, non-invasive machine learning (ML) approach to derive the arterial input function (AIF) for <sup>15</sup>O-water PET scans, a crucial but invasive step for measuring cerebral blood flow (CBF). By training a Gaussian process model on image-derived carotid artery signals, they demonstrate that their ML-derived AIF (MLIF) yields CBF estimates comparable to traditional blood sampling, successfully differentiating physiological changes induced by acetazolamide.
What Problem Does This Paper Address?
Accurately measuring cerebral blood flow (CBF) is vital for understanding brain function and diagnosing neurological conditions. Dynamic <sup>15</sup>O-water PET imaging, combined with tracer kinetic modeling, is a gold standard for this. However, a critical component of this modeling is the arterial input function (AIF), which represents the concentration of the tracer in the arterial blood over time. Traditionally, the AIF is obtained through continuous arterial blood sampling, typically from the radial artery. This process is invasive, time-consuming, and can be uncomfortable for patients, potentially discouraging participation in research. Furthermore, blood sampled from the arm requires corrections for dispersion and delay to accurately reflect the blood entering the brain. Image-derived input functions (IDIFs) have been proposed as an alternative, but they often suffer from partial volume effects and image noise, requiring complex correction methods. This paper aims to overcome these limitations by developing a non-invasive ML-based method to predict the AIF directly from PET image data.
Methods
The core of this work is a machine learning approach using Gaussian Processes (GP) for AIF prediction. The authors trained a GP model to predict the AIF (referred to as MLIF) based on time-activity curves extracted from image-derived regions of interest (ROIs) in the carotid arteries.
Here's a breakdown of the key methodological steps:
- Image-Derived Input Functions (IDIFs): Three different IDIFs were generated from the carotid arteries of PET scans. This involved a multi-VOI thresholding method to identify the 10, 100, and 1000 highest-intensity voxels within a defined search volume. These were labeled IDIF<sub>10</sub>, IDIF<sub>100</sub>, and IDIF<sub>1000</sub>, representing varying degrees of potential partial volume effects.
- Gaussian Process (GP) Regression: GPs are a non-parametric Bayesian regression method that predicts not only the mean function but also its variance, providing an uncertainty measure. This makes them suitable for data-driven function estimation, especially with potentially sparse or noisy data.
- Training and Testing Strategy: The study employed a leave-one-out cross-validation strategy. For each subject, one scan was held out as the test set, and the remaining scans were used for training the GP model. This was done for both baseline and acetazolamide scans separately (cases 1 and 2). A third approach (case 2) involved training a model on all baseline scans and testing it on acetazolamide scans to simulate a scenario where a model trained on healthy subjects is applied to a potentially altered physiological state.
- Kinetic Modeling for CBF: Once the AIF (either true AIF or MLIF) was obtained, whole-brain grey matter CBF was calculated using a standard single-tissue compartment model. This model relates the tracer concentration in the tissue to the AIF and kinetic parameters.
- Data Normalization: The IDIF curves were normalized to improve the convergence of the GP models. This involved scaling each curve by its highest peak value.
Datasets
The study utilized data from 25 subjects who underwent two 10-minute dynamic <sup>15</sup>O-water brain PET scans. These scans were acquired under two conditions:
- Baseline: A scan performed before any intervention.
- Acetazolamide: A scan performed 15-30 minutes after intravenous administration of acetazolamide, a drug that dilates blood vessels and increases cerebral blood flow.
Continuous arterial blood sampling was performed concurrently with the PET scans to obtain the ground truth AIF. The data included subjects with multiple sclerosis (MS) and healthy volunteers, though the study did not differentiate between these groups for the MLIF evaluation.
Key Insights and Ideas
- Non-invasive AIF is Feasible: The most significant insight is that a machine learning model, trained on image-derived carotid artery signals, can accurately predict the arterial input function for <sup>15</sup>O-water PET, eliminating the need for invasive arterial blood sampling.
- Carotid Arteries as a Proxy: The study demonstrates that signals from the carotid arteries, readily accessible in brain PET scans, can serve as effective input for ML models to predict the AIF.
- GP Models Capture Physiological Changes: The MLIF model successfully captured the physiological changes in blood flow induced by acetazolamide, indicating its ability to generalize to different physiological states.
- Shape Matters: The study highlights that the shape of the input function (IDIF) can significantly influence CBF estimates, especially when comparing different physiological states. The AIF from the wrist did not reflect the changes seen in the carotid artery IDIF after acetazolamide, suggesting that local brain input functions are more sensitive to these changes.
- Potential for Clinical Translation: The MLIF method offers a simplified and non-invasive alternative, potentially making CBF measurements more accessible and less burdensome for patients.
Key Results
- Similar AIF and MLIF Curves: The predicted MLIF curves were found to be similar to the AIF curves obtained from blood sampling.
- Highly Correlated CBF Estimates: Whole-brain grey matter CBF values estimated using MLIF were highly correlated with those estimated using the AIF (r<sup>2</sup> > 0.9).
- Minimal Bias: Bland-Altman analysis showed a prediction bias close to zero between AIF-based and MLIF-based CBF estimates.
- Successful Differentiation of Acetazolamide Effect: The MLIF method accurately predicted an increase in CBF after acetazolamide administration, similar to the AIF-based method. The mean CBF increase was approximately 37% for AIF and 29% for MLIF (case 1).
- Model Generalization: While training and testing on the same scans yielded highly similar results, training on baseline data and testing on acetazolamide data (case 2) showed a reduced but still significant CBF increase (17%), indicating the model's ability to generalize, albeit with some differences.
- Fractional Arterial Blood Volume (V<sub>A</sub>) Near Zero: The fractional arterial blood volume in the tissue was found to be near zero for all scans, consistent with previous findings.
Limitations and Future Work
- Limited Training Data: The GP models were trained on a relatively small number of samples (up to 22), which could limit generalization to significantly dissimilar data (domain shift).
- Outlier Identification: Four outliers were identified based on abnormal regression slopes between AIF and MLIF. The study suggests these might be due to AIF abnormalities rather than MLIF issues, but further quality control methods for MLIF predictions are needed.
- Heterogeneous Data: The dataset included both healthy volunteers and MS patients, and data from two different scanners, which introduced heterogeneity that could have affected results. However, the limited sample size prevented detailed investigation of these effects.
- Focus on Whole-Brain CBF: The evaluation primarily focused on whole-brain grey matter CBF. Future work should investigate the performance of the MLIF method for smaller brain regions.
- Test-Retest Variability: The test-retest variability of the MLIF method needs to be investigated and compared to that of blood AIF.
- Generalizability to Other Conditions: The study suggests investigating the MLIF method's sensitivity to other diseases, such as those involving carotid stenosis, which can alter AIF shape.
- Tracer and Scanner Specificity: The MLIF models are trained for specific tracers and imaging systems. Future work should explore adapting the method to other tracers and potentially incorporating metabolite correction.
Summary updated