Deep learning-guided attenuation and scatter correction without using anatomical images in brain PET/MRI

Attenuation correction (AC) is essential component for quantitative PET imaging. However, in PET/MR imaging and dedicated brain PET devices, the attenuation map either suffers from a number of limitations or is not readily available in the absence of CT or transmission scan. To tackle this issue, a deep convolutional neural networks is proposed to perform joint attenuation and scatter correction in the image domain on the non-attenuation corrected PET images (PET-nonAC). The deep con-volutional neural network used in this work benefits from dilated convolutions and residual connections to establish an end-to-end PET attenuation correction (PET-DirAC). For the training phase, data of 30 patients who underwent brain ¹⁸F-FDG PET/CT scans were used to generate reference PET-CTAC and PET-nonAC images. A five-fold cross-validation scheme was used for training/evaluation of the proposed algorithm. The quantitative accuracy of the proposed method was evaluated against the commercial segmentation-based method (2-class AC map referred to as MRAC). For quantitative analysis, tracer uptake estimated from PET-DirAC and PET-MRAC was compared to PET-CTAC. The relative SUV bias was calculated for bone, soft-tissue, air cavities and the entire head, separately. The proposed approach resulted in a mean relative absolute error (MRAE) of 4.1±7.5% and 5.8±10.4% for the entire head and bone regions, respectively. Conversely, MRAC led to a MRAE of 8.1±10.2% and 17.2±6.1% for these two regions, respectively. A mean SUV difference of 0.3±0.6 was achieved when using the direct method (DirAC) while the MRAC approach led to a mean SUV difference of -0.5±0.7. The quantitative analysis demonstrated the superior performance of the proposed deep learning-based AC approach over MRI segmentation-based method. The proposed approach seems promising to improve the quantitative accuracy of PET/MRI without the need for concurrent anatomical imaging.