Impact of deep learning-based MRI reconstruction on image quality and apparent diffusion coefficient measurements in patients with intracranial tumors: a prospective intra-individual study

Background: Deep learning-based magnetic resonance imaging (MRI) reconstruction has shown promise in improving image quality and reducing scan time. However, its impact on diffusion-weighted imaging (DWI) and quantitative parameters such as apparent diffusion coefficient (ADC) in patients with intracranial tumors remains unclear. This study aims to investigate the impact of deep learning-based MRI reconstruction on the quality of diffusion-weighted images and the measurements of lesion ADC in patients with intracranial tumors. Additionally, it assessed the potential influence of this reconstruction method on the consistency of ADC-based quantitative analysis.

Methods: This single-center, prospective intra-individual study involved 58 consecutive adults diagnosed with intracranial space-occupying lesions through imaging and/or pathology. These patients underwent brain MRI between January and May 2025, and all received standard single-b-value DWI (b=1,000 s/mm²). The raw DWI data were reconstructed using two methods: a conventional algorithm and a vendor-supplied deep learning reconstruction (DLR) algorithm based on convolutional neural networks. An experienced neuroradiologist manually placed regions of interest (ROIs) on the slice containing the lesion, including the lesion itself, the surrounding perilesional tissue, and the background. For each ROI, signal intensity (SI) and its standard deviation (SD) were measured, along with the lesion’s ADC values. The image quality metrics and ADC values obtained from the two reconstruction methods were compared using paired t-tests.

Results: Compared with conventional reconstruction, DLR significantly reduced background SI (29.40±5.05 vs. 34.95±4.99; t=13.948, P<0.001) and background noise (14.63±3.50 vs. 20.61±4.17; t=14.721, P<0.001), corresponding to approximately a 30% noise reduction. Lesion SI was slightly but significantly higher with DLR (1,786.42±878.56 vs. 1,782.41±878.79; t=2.863, P=0.006), while SI in perilesional tissue exhibited no significant difference (P=0.259). DLR also demonstrated significantly lower lesion ADC values than conventional reconstruction [(1,247.10±567.37)×10⁻⁶vs. (1,324.41±617.97)×10⁻⁶ mm²/s, t=2.974, P=0.004], with an average decrease of about 6%.

Conclusions: Under a conventional single-b-value DWI protocol (b=1,000 s/mm²), deep learning-based MRI reconstruction improved image quality in patients with intracranial tumors by reducing background noise and slightly increasing lesion signal. However, it was also associated with a modest but systematic decrease in lesion ADC values, suggesting that quantitative diffusion metrics may be affected to some extent. Further validation is needed to determine the clinical significance of this ADC shift and to assess the potential utility of DLR in more demanding diffusion imaging settings.