Faster musculoskeletal magnetic resonance imaging (MRI) can reduce examination burden, but shorter scans often affect resolution and image signal. Deep learning reconstruction (DLR) may help address that trade-off in routine joint imaging. A recent original investigation published in European Radiology Experimental evaluated DLR in 1.5-T knee, shoulder, ankle and hip MRI. Higher-resolution DLR sequences maintained signal-to-noise ratio (SNR) and contrast-to-noise ratio (CNR), while improving visibility of key musculoskeletal structures.
Higher Resolution Across Routine Joint Imaging
The retrospective cohort included 39 adults who underwent clinical musculoskeletal MRI on a 1.5-T scanner at a private clinic. Each person had paired imaging during the same examination session. One sequence used routine standard resolution, while the paired sequence used higher resolution with DLR. The examinations covered the knee, shoulder, ankle and hip. Included contrasts were T1-weighted fast spin-echo, proton density-weighted fast spin-echo with fat saturation and three-dimensional T2-weighted fast advanced spin-echo. Adults were included when paired eligible sequences were available. Image pairs were excluded when one acquisition had severe motion or prosthesis artefact, suboptimal fat saturation or a contrast outside the inclusion criteria. Forty-two patients were initially evaluated. Three were not included because of movement artefacts in one standard-resolution acquisition or sequence contrasts outside the inclusion criteria. The final cohort included 22 females and 17 males.
Must Read: Dixon Fat-Water MRI in Musculoskeletal Imaging
The higher-resolution DLR protocol produced smaller pixel and voxel sizes across all joints. Acquisition time decreased for the shoulder, hip and ankle. The knee showed a small increase because timing changed differently across the individual sequences. The DLR method worked directly on the MRI console and reconstructed raw data immediately after acquisition, without extra external processing. The paired approach came from the early clinical implementation of DLR, when conventional and DLR sequences were acquired together for radiologist familiarisation. DLR sequences were added according to scheduling feasibility, rather than patient preselection.
Visibility Gains Across Structures
Three radiologists rated structure visibility with a five-point Likert scale. One experienced reader assessed the images independently, while two junior readers gave a consensus assessment. The rating covered tendons, fibrocartilages, cartilages, ligaments, bone marrow and interfaces. Fibrocartilage was not assessed in ankle examinations because the ankle does not contain structures equivalent to the menisci or labrum. The scale ranged from poor visibility to excellent delineation. A fourth reader measured apparent SNR and apparent CNR in bone marrow and muscle using circular regions of interest.
Agreement between readers was good for standard-resolution sequences and higher for DLR sequences. Higher-resolution DLR images reached good to very good agreement, with the strongest agreement seen for cartilage. Visibility scores improved with DLR for every assessed structure. Higher or similar scores appeared regardless of the joint examined and sequence contrast used. Most comparisons showed a consistent pattern across joints and sequence contrasts. In the smaller number of comparisons where the size of the benefit varied, DLR images remained either equal to or better than standard-resolution images. No musculoskeletal structure received worse ratings after DLR reconstruction. The qualitative gains therefore did not depend on a single joint. The pattern covered the main structures involved in musculoskeletal assessment, including tendons, cartilage, ligaments, bone marrow and tissue interfaces.
Stable Quantitative Measures and Practical Limits
Apparent SNR in muscle, apparent SNR in bone marrow and apparent CNR between muscle and bone marrow did not differ significantly between standard-resolution and higher-resolution DLR sequences. These measures remained stable despite smaller voxel sizes and shorter acquisition times in several joint protocols. The absence of significant interactions with joint or sequence also supports a consistent quantitative pattern. Different coils were used for different anatomical areas, but SNR and CNR did not significantly differ across joints in either standard-resolution or DLR images. The quantitative stability therefore accompanied the qualitative gains.
The image examples show practical changes in anatomical detail. In knee imaging, an osteochondral lesion of the medial femoral condyle had clearer margins on proton density-weighted fat-saturated DLR images. A posterior horn medial meniscus lesion with subchondral bone oedema was better defined on T1-weighted DLR images. In the shoulder, supraspinatus tendon calcification showed clearer shape and margins. In the ankle and hip-related examples, ligaments and the hamstring tendon origin appeared clearer. Some pathological findings were better evaluable, while some normal anatomical details were better appreciated. Limitations include the small sample size, retrospective single-centre design, absence of detailed clinical indications, no formal diagnostic confidence measure and no assessment in patients with implants or prostheses. Prospective multicentre work would help confirm generalisability and clinical applicability.
Higher-resolution DLR in 1.5-T musculoskeletal MRI improved visibility of tendons, fibrocartilages, cartilages, ligaments, bone marrow and interfaces while preserving apparent SNR and CNR. Acquisition time decreased in several joint protocols and stayed broadly comparable in the knee. The results support on-scanner DLR as a way to increase spatial resolution without reducing core quantitative image quality measures within the tested protocol and in routine imaging. The same pattern appeared across the evaluated joints and sequence contrasts, although broader prospective evaluation remains necessary before wider applicability can be confirmed overall.
Source: European Radiology Experimental
Image Credit: iStock
References:
Porta M, Agresti G, Laganà MM et al. (2026) Enhancing resolution and image quality in musculoskeletal MRI using deep learning reconstruction. Eur Radiol Exp; 10, 78.
