Computed Tomography slice interpolation in the longitudinal direction based on deep learning techniques: To reduce...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Large slice thickness in Computed Tomography (CT) data leads to insufficient longitudinal information, degrading diagnostic quality.
• Existing solutions like high-resolution CT (HRCT) increase radiation dose, while linear interpolation introduces artifacts.

Purpose of the Study:

• To develop a deep learning approach for reconstructing densely sliced CT data from sparse data without increasing radiation dose.
• To improve the quality of CT-based diagnosis by addressing information insufficiency in the longitudinal direction.

Main Methods:

• A U-net architecture was employed to reconstruct CT images from neighboring sparse slices.
• A parallel U-net architecture was proposed to ensure independent reconstruction of slices, preventing mutual influence.
• A range-clip technique was introduced to enhance reconstruction quality for specific organs, like the liver, by adjusting the training data range.

Main Results:

• The parallel U-net architecture reduced the mean absolute error of CT values in reconstructed slices by 22.05% compared to linear interpolation.
• Artifacts around organ boundaries were reduced.
• The range-clip algorithm further improved reconstruction accuracy for the liver (15.12%), left kidney (11.04%), right kidney (10.94%), and stomach (10.63%).
• The proposed parallel U-net architecture demonstrated superiority over the original U-net method.

Conclusions:

• The proposed deep learning method effectively reconstructs high-quality, densely sliced CT data from sparse inputs without dose increase.
• The parallel U-net architecture and range-clip technique offer significant improvements in accuracy and artifact reduction for CT image reconstruction.
• This approach holds promise for enhancing CT-based diagnostic capabilities.