A Deep Learning Model for Brain PET Dual Tracer Separation based on Clinical Dataset include 18F-FDG, 18F-Florbetapir and 18F-Flutemetamol PET rradiotracer

Title:

A Transformer Based Deep Learning Model for Signal Separation in Dual Tracer Brain PET Imaging

Purpose/Background:

The potential of amyloid PET radiotracers, such as 18F-Florbetapir (FBP) and 18F-Flutemetamol (FMM), in the non-invasive detection of amyloid plaques, a key pathological feature in Alzheimer's disease (AD), has been demonstrated in previous studies. Detectable presence of amyloid plaques can precede noticeable cognitive decline by approximately 10–15 years, underscoring the value of these tracers in early AD diagnosis. Amyloid-PET imaging provides a measure of cerebral perfusion and is closely associated with neural dysfunction, suggesting its promise as a neurodegeneration biomarker. On the other hand, 18F-FDG PET estimates cerebral glucose metabolism as a well-established marker linked to synaptic dysfunction and neurodegeneration levels. Clinical investigations highlight that amyloid-PET and FDG-PET offer complementary insights for both clinical diagnosis and prognostic assessment of patients. Performing separate scans for both tracers in different sessions presents challenges, including increased patient discomfort, radiation exposure, scanning duration, and cost. To address this concern, we propose a scenario where a patient undergoes a single scan with simultaneous injection of both FBP/FMM and FDG, obtaining a dual-tracer image. Subsequently, a deep learning-based model is utilized to discriminate between the two tracers (FBP/FMM and FDG) from the combined image derived by summation.

Methods:

A clinical dataset consisting of 120 PET scans with FDG and FBP/FMM was incorporated. The standard late acquisition for amyloid-PET imaging was conducted at 50 minutes post-injection for FBP and at 90 minutes for FMM. To obtain intensity-normalized PET images for all radiotracers, the uptake was normalized to the mean value of the pons and cerebellar vermis combined, serving as a reference region. All images underwent registration to the MNI (Montreal Neurological Institute) space. FDG and FBP/FMM images were combined (summed) to create a simulated dual-tracer image for each subject. For the separation of FDG and amyloid signals from the combined image, a transformer model named Swin UNEtTRansformers (SwinUNETR) was employed. Training involved a fivefold cross-validation approach. Qualitative and quantitative analyses were carried out by calculating bias (image1 - image2), relative error ((image1 - image2) / image2), and relative absolute error (|(image1 - image2)| / image2) for the synthesized FDG and FBP/FMM PET images. Subsequently, 166 regions were delineated according to the automated anatomical labeling atlas 3 (AAL3), followed by region-specific Bland & Altman analysis for the synthesized FDG and FBP/FMM PET images.

Results:

Upon visual inspection, it was observed that the generated FDG and FBP/FMM images demonstrated comparable activity distributions to the reference images across all cases, regardless of disease stage or severity (Figure 1). Analysis of the difference map revealed a minimal bias range (-0.2 to +0.2 SUVR) for both FDG and FBP/FMM images across the whole brain. Bland & Altman graphs indicated mean differences of 0.02 and -0.07 SUVR, with confidence intervals ranging from -0.21 to 0.25 and -0.41 to 0.27 for FDG and FBP/FMM PET images, respectively (Figure 2). Each point on these graphs represents a specific region from an individual subject. Table 1 displays a relative error of -1.51 ± 1.87% for FDG and -0.01 ± 2.32% for FBP/FMM PET images for the whole brain.

Conclusion:

In this work, an end-to-end transformer-based method for estimating/separating the FDG and amyloid PET images from a single dual-tracer PET image was presented and evaluated. Our model generates FDG and Amyloid images with very similar radioactivity distribution to the actual FDG and amyloid PET images. Our model may facilitate the implementation of dual-tracer imaging in research and clinical settings by providing more comfort to patients, increasing scanner throughput, and providing a higher level of confidence for the decision-making process.