Deep-learning-based image quality enhancement of CT-like MR imaging in patients with suspected traumatic shoulder injury

Purpose: To evaluate the diagnostic performance of CT-like MR images reconstructed with an algorithm combining compressed sense (CS) with deep learning (DL) in patients with suspected osseous shoulder injury compared to conventional CS-reconstructed images. Methods: Thirty-two patients (12 women, mean age 46 ± 14.9 years) with suspected traumatic shoulder injury were prospectively enrolled into the study. All patients received MR imaging of the shoulder, including a CT-like 3D T1-weighted gradient-echo (T1 GRE) sequence and in case of suspected fracture a conventional CT. An automated DL-based algorithm, combining CS and DL (CS DL) was used to reconstruct images of the same k-space data as used for CS reconstructions. Two musculoskeletal radiologists assessed the images for osseous pathologies, image quality and visibility of anatomical landmarks using a 5-point Likert scale. Moreover, signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) were calculated. Results: Compared to CT, all acute fractures (n = 23) and osseous pathologies were detected accurately on the CS only and CS DL images with almost perfect agreement between the CS DL and CS only images (κ 0.95 (95 %confidence interval 0.82–1.00). Image quality as well as the visibility of the fracture lines, bone fragments and glenoid borders were overall rated significantly higher for the CS DL reconstructions than the CS only images (CS DL range 3.7–4.9 and CS only range 3.2–3.8, P = 0.01–0.04). Significantly higher SNR and CNR values were observed for the CS DL reconstructions (P = 0.02–0.03). Conclusion: Evaluation of traumatic shoulder pathologies is feasible using a DL-based algorithm for reconstruction of high-resolution CT-like MR imaging.