Learning with privileged knowledge of multiple kernels via joint prediction for CT Kernel conversion.

Area of Science:

• Medical Imaging
• Computer Vision
• Machine Learning

Background:

• Current CT kernel conversion models typically use single-source kernel inputs, limiting performance.
• Leveraging information from multiple kernels during training can significantly improve conversion accuracy.
• Clinical settings often restrict acquired data to a single kernel, necessitating robust single-input models.

Purpose of the Study:

• To develop a privileged knowledge learning framework for CT kernel conversion.
• To utilize auxiliary kernel information from training data to guide single-source kernel conversion.
• To improve the conversion of CT images from a specific source kernel to a target kernel using a joint prediction (JP) task.

Main Methods:

• An ensemble of kernel-specific (KS) networks (KSNets) was constructed for target kernel conversion.
• A joint prediction (JP) task was implemented for regularization and enhanced feature representation learning.
• A cross-shaped window-based attention mechanism was employed within the JP task to focus on relevant features and mitigate noise.

Main Results:

• The privileged knowledge learning framework was evaluated on a diverse clinical dataset (Siemens, GE, Philips).
• Experimental results confirmed the framework's effectiveness in enhancing CT kernel conversion.
• The method demonstrated improvements in detail and structure representations for converted images.

Conclusions:

• The proposed privileged knowledge learning framework significantly improves CT kernel conversion outcomes.
• Enhanced kernel conversion contributes to improved diagnostic accuracy in medical imaging.
• The framework advances quantitative measurement research through more reliable comparative analyses.