Deep learning reconstruction for zero echo time lung magnetic resonance imaging: impact on image quality and lesion detection.

ABSTRACT

AIMS: This study aimed to examine the impact of deep-learning reconstruction (DLR) on zero echo time (ZTE) lung MRI. MATERIALS AND METHODS: Fifty-nine patients who underwent both chest CT and ZTE lung magnetic resonance imaging (MRI) were enrolled. Noise reduction in ZTE lung MRI was compared using various DLR intensities (DLR-M, DLR-H) and conventional image filtering techniques (NF1 âˆ¼ NF4). The normalized noise power spectrum (NPS) was analysed through phantom experiments. Image sharpness was evaluated using a blur metric. We compared subjective image quality and the detection of sub-centimetre nodules and emphysema between the original and noise-reduced images. Statistical analyses included the Wilcoxon signed-rank and McNemar's tests, with inter-reader agreement assessed via Kappa coefficients. RESULTS: NPS peaks were lower in NF1 through NF4, DLR-M, and DLR-H compared to the original images. While the average spatial frequency of the NPS shifted towards lower frequencies with increasing NF levels, it remained unchanged with DLR. Blur metric values of NF1∼NF4 were significantly higher than those of the original images (p<0.008). However, there were no significant differences in blur metric values between DLR-M, DLR-H, and the original images. Image quality was rated highest for DLR-H, with a statistically significant improvement over the original (p<0.05). DLR-H showed higher diagnostic confidence for detecting sub-centimetre nodules than the original images. DLR-H showed higher diagnostic performance than the original for detecting emphysema. CONCLUSIONS: DLR can improve ZTE lung MRI quality while preserving image texture and sharpness, thereby enhancing the potential of ZTE for evaluating pulmonary parenchymal disease.