Predicting CT Image From MRI Data Through Feature Matching With Learned Nonlinear Local Descriptors.

Area of Science:

• Medical Imaging
• Radiology
• Computational Imaging

Background:

• Accurate attenuation correction in PET/MRI and MR-based radiation therapy planning are hindered by a lack of high-energy photon attenuation data.
• Magnetic Resonance Imaging (MRI) lacks direct electron density information crucial for accurate attenuation correction.

Purpose of the Study:

• To develop a novel method for predicting pseudo Computed Tomography (pCT) images from MRI data.
• To improve attenuation correction in PET/MRI and dose planning in MR-based radiation therapy.

Main Methods:

• Utilized learned nonlinear local descriptors and feature matching to predict pCT images from T1- and T2-weighted MRI.
• Employed explicit feature maps, low-rank approximation, and supervised manifold regularization for descriptor enhancement.
• Applied k-nearest neighbor regression for pCT patch estimation based on nearest neighbors in training data.

Main Results:

• Achieved a mean absolute error (MAE) of 75.25 ± 18.05 Hounsfield units for pCT prediction.
• Obtained a peak signal-to-noise ratio (PSNR) of 30.87 ± 1.15 dB.
• Demonstrated competitive performance in PET attenuation correction (1.56 ± 0.5% relative MAE) and dose planning (0.055 ± 0.107% relative structure volume difference), outperforming four state-of-the-art methods.

Conclusions:

• The proposed method effectively generates accurate pCT images from MRI data.
• This approach offers a viable solution for enhancing PET/MRI attenuation correction and MR-based radiation therapy planning.
• The technique shows significant potential for advancing hybrid imaging and radiotherapy applications.