PT  - JOURNAL ARTICLE
AU  - Izadi, Saeed
AU  - Shiri, Isaac
AU  - Uribe, Carlos
AU  - Geramifar, Parham
AU  - Zaidi, Habib
AU  - Rahmim, Arman
AU  - Hamarneh, Ghassan
TI  - <strong>Enhanced Direct Joint Attenuation and Scatter Correction of Whole-Body PET Images via Context-Aware Deep Networks</strong>
DP  - 2022 Aug 01
TA  - Journal of Nuclear Medicine
PG  - 2398--2398
VI  - 63
IP  - supplement 2
4099  - http://jnm.snmjournals.org/content/63/supplement_2/2398.short
4100  - http://jnm.snmjournals.org/content/63/supplement_2/2398.full
SO  - J Nucl Med2022 Aug 01; 63
AB  - 2398 Introduction: In positron emission tomography (PET), attenuation and scatter corrections is necessary steps towards accurate quantitative reconstruction of the radiopharmaceutical distribution. A known characteristic of PET imaging is to have varying tracer uptakes for various patients and/or anatomical regions. However, existing deep learning-based algorithms utilize a fixed model across different subjects and/or anatomical regions during inference, which could result in spurious outputs. In this work, we present a novel deep learning-based framework for direct reconstruction of attenuation and scatter corrected PET from non-attenuation-corrected images in absence of structural information, which is able to efficienly deal with the inter-subject and intra-subject uptake variations in PET imaging.Methods: We propose a novel model to perform subject- and region-specific filtering through modulating the convolution kernels in accordance to the contextual coherency within the neighboring slices. This way, the context-aware convolution can guide the composition of intermediate features in favor of regressing input-conditioned and/or region-specific tracer uptakes. In particular, the neighboring slices are fed into a shared sub-network to extract the region-wise information, which is subsequently used to modify the convolution kernels in an adaptive manner. Our proposed model differs from existing methods by (1) the 3D contextual information within neighboring slices is used to adaptively modulate the kernels in specific layers and (2) the network is still in 2D manner and therefore enjoys accessing to large training instances as well as light computation cost at inference. Whole-body 18F-FDG PET/CT scan data of 910 subjects were employed in this study for training and validation purposes.Results: The performance evaluation of our proposed context-aware network (CA-DAC) includes validation against the reference PET-CT ground truth and comparison to a conventional 2.5D UNet (UN-DAC) over a held-out test cohort consisting of 112 subjects across whole-body and 6 anatomical regions. In the whole-body, the average relative error (RE) were 2.43% ± 2.94 and -2.11% ± 2.73 for UN-DAC and CA-DAC, respectively. Also, the average absolute relative error (ARE) reduced from 14.79% ± 2.37 to 13.96% ± 2.32 for UN-DAC and CA-DAC, respectively. We also observed that CA-DAC outperformed UN-DAC by 0.71 (dB) averaged over all regions in terms of peak-signal-to-noise-ratio (PSNR). Likewise, the structural similarity (SSIM) scores averaged over all regions were improved from 0.9406 to 0.9465 using CA-DAC over UN-DAC in our evaluation set. The coefficient of determination, denoted by R2 was further used to quantify the goodness of fit within the joint histograms. In particular, R2 for CA-DAC and UN-DAC was recorded as 0.982 and 0.963, respectively, indicating a better fit using the former, i.e. proposed method. Moreover, we performed a linear regression analysis by reporting the slope and intercept of the line fitted over the non-zero bins of the joint histograms. UN-DAC yielded 0.81 and 1.30 for the slope and intercepts, respectively, while CA-DAC results in 0.88 and 0.57 demonstrating an increase of 0.07 for the slope (closer to 1.0) and a decrease of 0.73 for the intercept (closer to 0.0). Conclusions: In this work, we introduced CA-DAC to produce attenuation-corrected PET images without requiring any anatomical information during training and inference. The novelty of the proposed network is to take advantage of context-aware convolutions to modulate the convolution kernels based on the contextual information within neighboring slices along the axial dimension for every 2D input slice. This way, the network can effectively adapt itself to the inter- and intra-subject tracer uptake variations with negligible increase in model complexity.