Deep learning-based reconstruction for 5.0T magnetic resonance imaging (MRI) in nasopharyngeal carcinoma: comparison of image quality and diagnostic efficacy.

AIM: To investigate the effect of deep learning-based reconstruction (DLR) technology on image quality and diagnostic efficacy of 5T magnetic resonance imaging (MRI) in nasopharyngeal carcinoma (NPC). MATERIALS AND METHODS: This prospective study included 70 NPC patients who underwent 5T MRI examinations. The protocol included axial T2-weighted imaging (T2WI), axial T1-weighted imaging (T1WI), and axial and coronal contrast-enhanced T1WIs. Images of six gear levels (0-5) were reconstructed using the DLR technology. Two radiologists independently evaluated the visibility of lesions, boundary sharpness, artefact presence, and overall image quality using a 5-point Likert scale. Signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR) were also calculated for lesions and lateral pterygoid muscles (LPM). T-stage evaluation was performed for both conventional (level = 0) and DLR images, compared to clinical T-stage results. RESULTS: DLR images (levels 2-5) demonstrated significantly higher SNRs in lesions and LPM compared to conventional images (P < 0.001), with a maximum improvement of 94%. In axial T2WI, and axial and coronal contrast-enhanced T1WI DLR images (levels 2-5), the lesion-to-lateral pterygoid muscle CNR was significantly higher (P < 0.001), with a maximum improvement of 108%. Qualitative analysis revealed that DLR images (levels 2-5) were superior to conventional images (P < 0.05) across all subjective assessment dimensions except artefact reduction. Among the five reconstruction levels of DLR, level 3 yielded the highest overall image quality score. Additionally, the diagnostic performance of T-stage for DLR (level 3) and conventional images exhibited a comparable degree of consistency with the clinical T-stage results (κ = 0.771 and 0.796, respectively). CONCLUSION: DLR technology improves the quality of conventional nasopharyngeal MRI images without affecting the diagnostic accuracy of T-stage, offering potential clinical value.