The value of a deep learning image reconstruction algorithm for assessing vertebral compression fractures using dual-energy computed tomography.

PURPOSE

To evaluate the value of deep learning image reconstruction (DLIR) in improving image quality of virtual non-hydroxyapatite (VNHAP) and virtual monoenergetic images (VMIs), and radiologists' performance in detecting acute vertebral compression fractures (VCFs).

METHODS

VMIs at 70 keV and VNHAP images from 103 vertebrae (46 normal vertebra, 29 acute and 28 chronic VCFs) were reconstructed with four algorithms: adaptive statistical iterative reconstruction (AR50), DLIR-Low (DL), DLIR-Middle (DM), and DLIR-High (DH). Objective indexes including CT values for 70 keV VMIs and water density for VNHAP images, noise, signal-to-noise ratio [SNR], contrast-to-noise ratio [CNR] of all vertebral bodies, as well as intervertebral ratio [IVR] of acute and chronic VCFs. Subjective image quality scoring and VCFs diagnosis were independently performed by four radiologists. Employing MR examinations as the reference, the specificity, sensitivity, accuracy, and predictive metrics for detecting acute VCFs were assessed.

RESULTS

DLIR reduced image noise and improved SNR and CNR for all vertebra on both 70 keV VMIs and VNHAP images. DH-reconstructed VNHAP images had the highest IVR for acute VCFs and outperformed for all comparisons. The four-reader subjective scores on image quality: DH > DM > DL > AR50. For the differentiation between acute and chronic VCFs, the four readers with VNHAP images showed enhanced performance than those with 70 keV VMIs, while DLIR further improved the efficiency, with DH performed the best.

CONCLUSION

DLIR improved the quality of VMIs and VNHAP images, elevated the contrast of acute VCFs, and enhanced radiologists' performance in detecting acute VCFs. DH performed the best.