Intravoxel incoherent motion imaging with improved image quality by deep learning reconstruction for prediction of lymph-vascular space invasion in patients with cervical cancer.

PURPOSE: To evaluate the performance of intravoxel incoherent motion (IVIM) imaging with deep learning reconstruction (DLR) for prediction of lymph-vascular space invasion (LVSI) in cervical cancer (CC). METHODS: Patients clinically suspected of having CC who underwent uterine MRI were retrospectively collected. IVIM images were collected with b values of 0, 20, 50, 100, 150, 200, 400, 800, 1500 and 2000 s/mm2. Subjective image quality, signal-to-noise ratio (SNR), contrast-to-noise ratio (CNR), as well as apparent diffusion coefficient (ADC) values calculated at b = 0 and 800 s/mm² (ADC800), ADC values at b = 0 and 2000 s/mm² (ADC2000), and IVIM quantitative parameters were computed and compared between IVIM sequences with and without DLR. Quantitative parameters measured by the two sets of IVIM data were also compared between LVSI-positive and LVSI-negative groups. RESULTS: The study included 55 patients with 31/55 (56.36%) lesions identified as LVSI positive. IVIM images with DLR were associated with significantly higher image quality scores, SNR, and CNR than those without DLR. All quantitative parameters except ADC800 showed significant differences between the two data sets. For both IVIM with and without DLR, the ADC2000 and true diffusion coefficient (D) values were significantly different between LVSI-positive and LVSI-negative groups. The combination of ADC2000 and D measured from IVIM images with DLR yield the highest AUC of 0.809. CONCLUSION: Reconstruction of pelvic IVIM images with DLR significantly improved image quality. IVIM parameters derived from DLR images showed potential for non-invasive prediction of LVSI status in CC.