Imitation learning for improved 3D PET/MR attenuation correction.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Pseudo CT (pCT) image quality is typically assessed by intensity similarity to CT.
• This metric overlooks the primary goal in PET/MRI: accurate PET reconstruction using pCT as an attenuation map.
• Minimizing pCT-CT error does not guarantee optimal PET reconstruction.

Purpose of the Study:

• To develop a novel deep learning framework for synthesizing pCT images tailored for PET reconstruction in PET/MRI.
• To address the limitations of traditional intensity-wise similarity metrics for pCT assessment.
• To improve the accuracy of pseudo PET (pPET) reconstructions derived from pCT attenuation maps.

Main Methods:

• A multi-hypothesis deep learning framework using a convolutional neural network (CNN) was developed.
• The CNN synthesizes pCTs by minimizing a combined loss: pixel-wise error (pCT-CT) and a novel metric-loss.
• The metric-loss, defined by another CNN, specifically targets the reduction of consequent PET reconstruction errors.
• Training utilized a database of 20 3D MR/CT/PET brain image pairs.

Main Results:

• The proposed method significantly improved pCT accuracy, achieving a Mean Absolute Error of 110.98 HU ± 19.22 HU.
• This outperformed a baseline CNN (172.12 HU ± 19.61 HU) and a multi-atlas propagation approach (153.40 HU ± 18.68 HU).
• Consequently, PET reconstruction error was substantially reduced to 4.74% ± 1.52%, compared to 13.72% ± 2.48% (baseline) and 6.68% ± 2.06% (multi-atlas).

Conclusions:

• The novel deep learning framework effectively synthesizes accurate pCT images for PET/MRI.
• The method directly optimizes for improved PET reconstruction, surpassing conventional approaches.
• This represents a significant advancement in generating reliable attenuation maps for PET imaging in hybrid scanners.