Pseudo CT Estimation from MRI Using Patch-based Random Forest.

Area of Science:

• Medical Imaging
• Radiotherapy Physics
• Machine Learning in Medicine

Background:

• Computed Tomography (CT) simulators pose radiation risks in radiation therapy planning.
• Magnetic Resonance (MR) simulators offer a radiation-free alternative, but require accurate CT-equivalent data for treatment planning.

Purpose of the Study:

• To develop and validate a novel method for estimating pseudo CT images from MR images.
• To enable MRI-based radiation therapy planning and attenuation correction in PET/MRI scanners.

Main Methods:

• A patch-based random forest algorithm was employed for pseudo CT estimation.
• Patient-specific anatomical features were extracted from aligned training MR images as voxel signatures.
• Feature selection was used to identify robust and informative features for training the random forest model.

Main Results:

• The proposed method accurately generated pseudo CT images from MR images, as validated with human brain imaging data.
• Prediction accuracy was quantified using Peak Signal-to-Noise Ratio (PSNR) and Feature Similarity (FSIM) indexes.
• The technique demonstrated clinical feasibility for generating accurate pseudo CT data.

Conclusions:

• A novel patch-based random forest method for pseudo CT prediction from MR images has been developed.
• The method is clinically feasible and accurately predicts CT images, offering a radiation-free alternative for treatment planning.
• This technique is a valuable tool for MRI-based radiation therapy planning and PET/MRI attenuation correction.