A CT image feature space (CTIS) loss for restoration with deep learning-based methods.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Radiology

Background:

• Deep learning methods are prevalent in medical imaging for tasks like detection and segmentation.
• Loss functions significantly impact CT image restoration quality, affecting clinical diagnosis.
• Existing loss functions may not always yield optimal results for CT image restoration.

Purpose of the Study:

• To compare commonly used loss functions in deep learning-based CT image restoration.
• To propose a generalizable framework for loss functions in the feature space.
• To introduce novel loss functions, CTIS loss and RaW-LoFS, for improved CT image quality.

Main Methods:

• Developed a CT image feature space (CTIS) loss using an autoencoder trained on high-quality CT images.
• Proposed random-weight loss in the feature space (RaW-LoFS) for scenarios lacking high-quality data.
• Evaluated loss functions using post-reconstruction deep learning methods on the 2016 AAPM low-dose CT challenge dataset.

Main Results:

• CTIS loss and RaW-LoFS outperformed the widely used perceptual loss quantitatively and qualitatively.
• Subjective assessments by radiologists indicated superior visual quality for CTIS loss.
• A partially constrained CTIS loss was developed through channel analysis.

Conclusions:

• The proposed CTIS loss and RaW-LoFS achieve favorable image quality in CT restoration.
• The feature-space loss function framework demonstrates broad applicability to other tasks and fields.
• These novel loss functions hold promise for enhancing diagnostic accuracy in low-dose CT imaging.