The value of a deep learning image reconstruction algorithm for assessing vertebral compression fractures using...

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Virtual monoenergetic images (VMIs) and virtual non-hydroxyapatite (VNHAP) images are crucial for evaluating vertebral compression fractures (VCFs).
• Image reconstruction algorithms impact the diagnostic quality of these images.

Purpose of the Study:

• To assess the efficacy of deep learning image reconstruction (DLIR) in enhancing VMI and VNHAP image quality.
• To evaluate the impact of DLIR on radiologists' accuracy in detecting acute VCFs.

Main Methods:

• Reconstruction of 103 vertebrae using adaptive statistical iterative reconstruction (AR50) and three DLIR levels (DL, DM, DH).
• Objective assessment of image quality metrics (noise, SNR, CNR, IVR) and subjective scoring by four radiologists.
• Comparison of diagnostic performance for acute VCFs against MRI reference standard.

Main Results:

• DLIR algorithms reduced noise and improved SNR and CNR across all image types.
• The DLIR-High (DH) algorithm yielded the highest intervertebral ratio for acute VCFs and superior subjective image quality scores.
• Radiologists demonstrated improved VCF detection with VNHAP images, further enhanced by DLIR, particularly the DH algorithm.

Conclusions:

• Deep learning image reconstruction effectively improves VMI and VNHAP image quality.
• DLIR enhances the conspicuity of acute VCFs and improves diagnostic performance.
• The DLIR-High algorithm represents the optimal choice for VCF detection.