CT kernel conversions using convolutional neural net for super-resolution with simplified squeeze-and-excitation...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Computed tomography (CT) kernel variations impact diagnostic accuracy.
• Storing multiple CT kernel volumes presents significant storage and maintenance challenges.
• Developing efficient methods for CT kernel conversion is crucial for optimizing clinical workflows.

Purpose of the Study:

• To propose a novel method for converting CT image kernels using convolutional neural networks (CNNs).
• To address the limitations of storing and maintaining multiple CT kernel datasets.
• To enhance diagnostic accuracy by enabling flexible kernel conversion.

Main Methods:

• Utilized a super-resolution (SR) network with Squeeze-and-Excitation (SE) blocks for CT kernel conversion.
• Implemented both single-conversion and multi-conversion models, incorporating progressive learning (PL) and auxiliary losses.
• Evaluated conversion quality using root-mean-square-error (RMSE), structural similarity (SSIM) index, and mutual information (MI).

Main Results:

• The multi-conversion model demonstrated significantly superior image quality compared to conventional methods.
• Achieved high SSIM index values (e.g., 0.998 ± 0.001 for B10f-B30f conversion).
• The proposed deep learning approach effectively converted between various CT kernel types (e.g., smooth to sharp and vice versa).

Conclusions:

• Deep learning-based CT kernel conversion using SR networks offers a viable solution for managing CT data.
• The integration of SE blocks and progressive learning significantly enhances model performance.
• This method provides a promising approach for improving CT image quality and diagnostic capabilities.