Deep learning reconstruction for liver DWI: impact on image quality and ADC quantification

To assess the image quality, apparent diffusion coefficient (ADC) values, and lesion detection of diffusion-weighted imaging (DWI) with compressed sensing (CS) and model-based deep learning (DL) reconstruction (DL-DWI) by comparing with CS-DWI.

We retrospectively analyzed 188 patients (103 men and 85 women; mean age, 69.6 years) who underwent liver DWI between October 2024 and January 2025. DL-DWI incorporated Adaptive-CS-Net, a model-based deep learning architecture embedded within the CS reconstruction pipeline. Two radiologists independently performed quantitative and qualitative evaluations, including lesion conspicuity assessment.

Signal-to-noise ratio in DL-DWI was higher than that in CS-DWI for all organs (p < 0.05). Coefficient of variation (except for lateral segment and paraspinal muscle) and ADC values in DL-DWI were significantly lower than those in CS-DWI for all organs (p < 0.05). Contrast-to-noise ratio of all lesions (hepatocellular carcinoma, metastasis, and hemangioma; n = 155) was significantly larger in DL-DWI than in CS-DWI (p < 0.05). Image quality scores for the liver contours, distortion of the liver, image noise, artifacts, and overall image quality were significantly higher for DL-DWI than for CS-DWI (p < 0.05). Overall lesion conspicuity (n = 227) was significantly higher in DL-DWI than in CS-DWI (p < 0.001). Lesion conspicuity of small lesions (<10 mm, n = 72) and superficial lesions (within 1 cm from the liver surface, n = 75) in DL-DWI was significantly higher than in CS-DWI (p < 0.05).

Model-based DL reconstruction improved qualitative and quantitative image quality and lesion conspicuity, particularly for small and superficial lesions.