Unsupervised deep learning framework for data-driven gating in positron emission tomography

Abstract

Background: Physiological motion, such as respiratory motion, has become a limiting factor in the spatial resolution of positron emission tomography (PET) imaging as the resolution of PET detectors continue to improve. Motion-induced misregistration between PET and CT images can also cause attenuation correction artifacts. Respiratory gating can be used to freeze the motion and to reduce motion induced artifacts.

Purpose: In this study, we propose a robust data-driven approach using an unsupervised deep clustering network that employs an autoencoder (AE) to extract latent features for respiratory gating.

Methods: We first divide list-mode PET data into short-time frames. The short-time frame images are reconstructed without attenuation, scatter, or randoms correction to avoid attenuation mismatch artifacts and to reduce image reconstruction time. The deep AE is then trained using reconstructed short-time frame images to extract latent features for respiratory gating. No additional data are required for the AE training. K-means clustering is subsequently used to perform respiratory gating based on the latent features extracted by the deep AE. The effectiveness of our proposed Deep Clustering method was evaluated using physical phantom and real patient datasets. The performance was compared against phase gating based on an external signal (External) and image based principal component analysis (PCA) with K-means clustering (Image PCA).

Results: The proposed method produced gated images with higher contrast and sharper myocardium boundaries than those obtained using the External gating method and Image PCA. Quantitatively, the gated images generated by the proposed Deep Clustering method showed larger center of mass (COM) displacement and higher lesion contrast than those obtained using the other two methods.

Conclusions: The effectiveness of our proposed method was validated using physical phantom and real patient data. The results showed our proposed framework could provide superior gating than the conventional External method and Image PCA.

Keywords: data-driven; deep clustering; respiratory gating; unsupervised learning.

FIGURE 1.

(a) Autoencoder-based deep clustering network framework. (b) The schematic diagram of the AE with a self-attention gate. The number of features in each layer are indicated at the bottom of the graph.

FIGURE 2.

The overall diagram of our proposed unsupervised deep learning framework for data-driven gating.

FIGURE 3.

(a) Picture of the respiratory motion phantom and (b) the acquisition setup with the Anzai belt attached to the phantom.

FIGURE 4.

(a) Mechanical drawing of the rotating disk. There are 4 holes in the rotating disc, and radial distance of each hole to the rotating center is 40 mm for “A”, 30 mm for “B”, 20 mm for “C” and 25 mm for “D”. (b) Mechanical drawing of each hole. A hot source (inner diameter: 16mm) was filled using a smaller cylinder in the hole “A”, and the other 3 holes were left empty (the diameter of each hole is 23mm). The empty holes are filled with the background activity.

FIGURE 5.

Reconstruction of gated data using (a) External, (b) Image PCA and (c) Deep Clustering methods.

FIGURE 6.

Comparison between Image PCA and Deep Clustering methods. (a) Percentage gating accuracy. (b) Mean absolute phase shift relative to the gate centers. (c) Distribution of the count level over gates for the External, Image PCA and Deep Clustering methods.

FIGURE 7.

Gated PET reconstructions at end-inspiration phase from Deep Clustering method superimposed with (a) free-breathing helical CT and (b) phase matched CT.

FIGURE 8.

Sampled reconstruction of gated images using (a) External, (b) Image PCA, (c) Deep Clustering methods. The yellow line marks the top of the myocardium position in Gate 3 (the reference gate). The average FWHMs over 25 different locations of the intensity profile (indicated by the red line) through the (d) inferior myocardial wall were plotted as a function of gate number (solid lines). The dashed lines denote the mean FWHMs over all gates of the three methods.

FIGURE 9.

Sampled reconstruction of gated images using External, Image PCA, Deep Clustering methods. The red circle marks the lesion used for quantification in Figure 10.

FIGURE 10.

Averaged contrast for a pulmonary lesion (marked by the red circle in Fig. 9) as a function of the lesion ROI size (number of voxels) for the External, Image PCA, Deep Clustering methods.

FIGURE 11.

Comparison of the frequency spectrums of (a) the Anzai signal and (b,c) the latent features with the maximum energy inside the respiratory frequency range extracted by the AE (b) with and (c) without the self-attention module.