A two-dimensional feasibility study of deep learning-based feature detection and characterization directly from CT...

Area of Science:

• Medical Imaging
• Machine Learning in Radiology
• Computed Tomography (CT) Analysis

Background:

• Machine learning, particularly deep learning, is widely applied in computed tomography (CT) for image reconstruction and enhancement.
• Current CT analysis typically operates on reconstructed images, making it susceptible to artifacts like partial volume, beam hardening, and motion.

Purpose of the Study:

• To develop and evaluate machine learning methods for feature detection and analysis directly from CT projection data (sinogram domain).
• To enable blood vessel detection and characterization in the sinogram domain, avoiding reconstruction-induced artifacts.

Main Methods:

• Utilized a residual encoder-decoder convolutional neural network (REDCNN) to estimate sinogram domain vessel maps.
• Extracted vessel-only sinograms by estimating vessel centerlines and eliminating background noise.
• Employed a fully connected neural network to determine vessel lumen cross-sectional area from the processed sinogram data.

Main Results:

• Demonstrated the feasibility of CT analysis directly within the sinogram domain using simulations, phantom data, and clinical datasets.
• Achieved encouraging initial results for blood vessel detection and characterization without relying on reconstructed images.

Conclusions:

• Sinogram domain analysis presents a viable alternative for CT data analysis, potentially applicable to various clinical tasks.
• Further research is necessary to refine this sinogram domain approach for widespread clinical adoption.