Diffusion-weighted imaging (DWI) plays a central role in liver MRI because it can improve detection and characterisation of focal liver lesions and supports quantitative assessment through apparent diffusion coefficient (ADC) measurements. In routine practice, however, liver DWI can be challenged by motion from breathing, the heartbeat and gastrointestinal activity, and image quality can be further affected when using higher diffusion weightings. A prospective evaluation explored whether deep learning reconstruction (DLR) could deliver faster liver DWI while improving image quality and maintaining ADC-based discrimination between benign and malignant lesions. The work compared conventional respiratory-triggered DWI with a DLR approach designed to enhance reconstruction, reduce noise and sharpen detail.
Faster Respiratory-Triggered DWI Through Deep Learning Reconstruction
Consecutive patients with suspected liver disease underwent liver MRI and received both conventional DWI and DLR-reconstructed DWI. Imaging was performed at 3.0 T using respiratory-triggered fat-suppressed single-shot echoplanar DWI in the transverse plane, with diffusion gradients applied in three directions. The protocol used b-values of 0 and 800 s/mm².
The accelerated approach reduced acquisition time by lowering the number of averages while keeping the respiratory-triggering scheme consistent between methods. Under the reported setup, scan time decreased from roughly six minutes for conventional DWI to around three and a half minutes with DLR. The reconstruction framework combined compressed sensing with convolutional neural networks intended to handle k-space reconstruction and improve the final image through denoising and resolution enhancement. The practical intent was not only to shorten the examination but also to counteract the loss of signal and clarity that can accompany faster diffusion acquisitions.
Must Read: MRI-Based FLIS Supports Liver Function Assessment
Clearer Lesion Depiction with Improved Signal and Edge Definition
Image quality was assessed using both reader scoring and quantitative measurements. Two radiologists independently reviewed both DWI datasets in a randomised, blinded fashion, focusing on overall image quality, sharpness, artefacts and lesion conspicuity. In parallel, regions of interest in liver parenchyma and lesions supported measurements of signal-to-noise and contrast-to-noise, with a muscle-based reference used to represent background variation.
Across the patient group, DLR-reconstructed DWI scored higher for lesion conspicuity, vessel visibility and overall image quality. Artefact ratings did not show a meaningful disadvantage for the DLR approach under the same respiratory-triggering conditions. Quantitative analysis aligned with these impressions. DLR-reconstructed DWI produced higher liver and lesion signal-to-noise and improved contrast between lesions and background liver. Lesion boundary definition was also sharper, reflected by a shorter edge rise distance, indicating a more abrupt transition between lesion and surrounding tissue. A small number of lesions were excluded from quantitative assessment due to insufficient image quality, and reproducibility across readers for the quantitative metrics was reported as good to excellent.
ADC-Based Discrimination Maintained as Values Shift Lower With DLR
The evaluation included 193 patients, with malignant and benign focal liver lesions represented. Malignant lesions included hepatocellular carcinoma, liver metastases and cholangiocarcinoma, while benign findings included haemangioma, focal nodular hyperplasia, adenoma, abscess and cyst. Lesion status was determined using consensus interpretation supported by typical MRI features, clinical history, pathology and follow-up imaging when applicable.
ADC values were lower with DLR-reconstructed DWI for both malignant and benign lesions. The analysis indicated that lesion type influenced ADC values strongly, while the reconstruction method itself did not demonstrate a statistically meaningful effect across lesion categories, and there was no interaction between method and lesion type. Despite the ADC shift, diagnostic performance for differentiating benign from malignant lesions improved modestly with DLR reconstruction on receiver operating characteristic analysis, with a higher area under the curve than conventional DWI. In performance terms, DLR-reconstructed DWI showed higher sensitivity and slightly higher overall accuracy, while specificity was somewhat lower, reflecting a different balance between detecting malignancy and avoiding false positives. A subgroup analysis focusing on malignant lesions versus selected benign lesions also favoured DLR, with a higher area under the curve and similar sensitivity between methods.
DLR-reconstructed liver DWI, implemented with respiratory triggering and standard diffusion weighting, was associated with faster acquisition and improved image quality compared with conventional DWI in a prospective cohort. The approach reduced scan time by about half while improving lesion conspicuity, perceived sharpness and quantitative measures of signal, contrast and edge definition, without an apparent penalty in artefact ratings. Although ADC values shifted lower with DLR reconstruction in both benign and malignant lesions, the method preserved lesion-type separation and delivered a modest improvement in ADC-based discrimination performance. For imaging services balancing throughput, patient tolerance and diagnostic confidence, these findings frame DLR as a route to shorter liver DWI acquisitions while supporting clinically relevant lesion assessment.
Source: Insights into Imaging
Image Credit: iStock
