Prior image-guided cone-beam computed tomography augmentation from under-sampled projections using a convolutional...

Area of Science:

• Medical Imaging
• Radiology
• Computer-Aided Diagnosis

Background:

• Sparse-view cone-beam computed tomography (CBCT) reduces radiation dose but suffers from artifacts due to under-sampling.
• Existing deep learning models struggle with patient-specific details in CBCT reconstruction.
• Generalizable models that leverage patient-specific information for under-sampled image augmentation are needed.

Purpose of the Study:

• To develop a generalized deep learning model for prior image-guided sparse-view CBCT augmentation.
• To improve the quality of CBCT images reconstructed from highly under-sampled projections.
• To reduce imaging dose while maintaining diagnostic accuracy.

Main Methods:

• Proposed a merging-encoder convolutional neural network (MeCNN) for CBCT image enhancement.
• MeCNN extracts and merges multi-scale features from prior CT and under-sampled CBCT images.
• Model tested on simulated and clinical CBCT data; evaluated qualitatively and quantitatively.

Main Results:

• MeCNN reconstructed high-quality, CT-like CBCT images from 36 half-fan projections.
• Achieved significantly lower intensity errors, higher peak signal-to-noise ratio, and structural similarity compared to other methods.
• Significantly reduced CBCT-based tumor localization errors with near real-time augmentation.

Conclusions:

• The proposed prior-image guided method effectively reconstructs high-quality CBCT images from sparse views.
• This approach significantly reduces imaging dose and enhances the clinical utility of CBCT.
• MeCNN offers a promising solution for dose reduction in medical imaging.