Machine Learned Texture Prior From Full-Dose CT Database via Multi-Modality Feature Selection for Bayesian...

Area of Science:

• Medical Imaging
• Radiology
• Computer Vision

Background:

• Previous low-dose CT (LdCT) reconstruction methods relied on assumptions of linear relationships and availability of prior full-dose CT (FdCT) scans.
• These assumptions limited the clinical applicability of texture prior-based reconstruction techniques.

Purpose of the Study:

• To develop a database-assisted end-to-end LdCT reconstruction framework that overcomes limitations of previous methods.
• To eliminate the need for linear relationship assumptions and prior FdCT scans from the same subject.

Main Methods:

• A convolutional neural network (CNN) was used to create a deep learning texture prior, removing the linear relationship assumption.
• A multi-modality feature-based candidate selection model was developed to select appropriate FdCT priors from a database when same-subject FdCT scans are unavailable.
• Features included physiological factors, CT scan protocols, and a novel 'Lung Mark' feature reflecting z-axial anatomy.

Main Results:

• The selection model achieved 84% accuracy in selecting appropriate priors from a database of 1,470 images (49 subjects).
• Reconstructions using learned texture priors from the FdCT database were comparable to those obtained with subject-specific FdCT scans.
• The 'Lung Mark' feature demonstrated effectiveness in improving reconstruction quality.

Conclusions:

• The proposed framework enables the use of clinically relevant textures from an FdCT database for Bayesian reconstruction of LdCT scans.
• This approach enhances LdCT image quality and expands its clinical utility, even without prior FdCT data from the same patient.