This study enrolled a prospective cohort of 83 patients scheduled to undergo knee MRI examinations between June and July 2025. Following screening, 69 patients were ultimately included.The knee arthroscopy-MRI interval was limited to 10 days across all patients. The cohort included 38 males and 31 females, aged 13–61 years (mean age: 36.63 ± 11.82). All patients underwent 5.0T knee MRI on a clinical scanner (uMR Jupiter, United Imaging Healthcare, China). The acquired data were processed using a deep learning reconstruction (DLR) algorithm at three distinct intensity levels: weak, medium, and strong. Inclusion Criteria: ①Clinical diagnosis or suspected knee pathologies warranting MRI evaluation. ②Images fulfilled diagnostic quality criteria. Exclusion Criteria: ①Contraindications to MRI (e.g., cardiac pacemakers, metallic implants)(n = 0). ②Insufficient imaging sequences or absence of raw reconstruction data (n = 1). ③Severe cartilage degeneration precluded region-of-interest (ROI) sampling (n = 4). ④Positional shifts during scanning (n = 4). ⑤Severe motion artifacts (n = 5). Figure 1 illustrates the patient screening workflow.

MRI was performed using a 5.0T scanner equipped with a 24-channel knee coil. During the examination, all patients were positioned supine in a feet-first orientation. Sequences included T2-weighted fast spin-echo (T2-FSE), T1-weighted FSE (T1-FSE), and density-weighted (PD)-weighted acquisitions. Parameters are summarized in Table 1.

DLR resolves the longstanding MRI trilemma (scan time, SNR, spatial resolution) through adaptive neural network-driven optimization (16). Traditional approaches necessitate trade-offs between scan duration and resolution to optimize SNR, k-space noise refers to the inherent random signal fluctuations in the raw acquisition data, which manifest as a granular texture in the background or uniform tissue regions of the reconstructed image. Gibbs artifacts are oscillatory stripes of alternating brightness and darkness that appear along sharp tissue boundaries in the image, resulting from the truncation of high-frequency data in k-space, while k-space-related noise and Gibbs artifacts persist with conventional mathematical frameworks (17, 18). The DLR, named DeepRecon (United Imaging Healthcare, China), implements lightweight-optimized convolutional neural networks (CNNs) to mitigate noise and improve spatial resolution while preserving acquisition durations. The CNN model used in this study is an end-to-end reconstruction framework. The input consists of undersampled raw k-space data, and the output is the corresponding fully-sampled image. The network employs a dual-domain learning architecture that simultaneously optimizes image-domain fidelity and k-space data consistency. An embedded data consistency module ensures the reconstruction conforms to the actual physical measurements. The model is trained via supervised learning to directly learn a nonlinear mapping from accelerated acquisitions to high-quality images. Rather than relying on fixed mathematical priors, the CNN learns a direct, non-linear mapping from undersampled k-space to high-quality images. Its key advantage lies in adaptive, context-aware processing—simultaneously suppressing noise and mitigating artifacts such as Gibbs ringing—while preserving true anatomical edges through dual-domain consistency constraints. The novelty resides in dual-domain (image/k-space) fidelity constraints: Image domain, harnessing multicenter training data encompassing anatomical and contrast variations, DLR enables discriminative noise-signal separation while preserving anatomical fidelity (19, 20); k-space domain, the fidelity module enforces data consistency between reconstructed images and raw k-space data through loss function constraints. Iterative optimization and matrix fusion synergistically improve accuracy. Experimental results demonstrate 44% SNR enhancement vs. conventional methods while maintaining image uniformity (13, 21–22).

The study comprised two phases: standard knee MRI acquisitions followed by DLR application with varying reconstruction intensities (weak/medium/strong), the Weak, Medium, and Strong reconstruction levels, which are integrated presets of the DLR algorithm balancing denoising and detail preservation. All images for different levels were generated from the same raw k-space data in an offline, parallel reconstruction process to ensure a controlled comparison.

Two radiologists with extensive diagnostic experience (25 years and 21 years respectively) independently diagnosed each patient's cartilage, meniscus, bone, ligaments, and muscles (in 2D sequences Sagittal T2WI FSE, Sagittal PD FSE FS, Sagittal T1WI FSE, Coronal PD FSE FS) using the following grading system. Grading system:

Meniscus (medial/lateral): Normal (0), degeneration (1), tear (2).

Cartilage (patellofemoral/tibiofemoral): Normal (0), degeneration (1), injury (2).

Ligaments (ACL/PCL): Normal (0), injury (1), tear (2).

Bone: Normal (0), bone marrow edema (1), cyst (2).

Muscle: Normal (0), edema (1), tear (2).

Statistical analysis was conducted with statistical software (SPSS, v. 27.0, R software). This study employed a combined quantitative and qualitative approach to evaluate the impact of deep learning reconstruction (DLR) on knee MRI image quality and diagnostic efficacy. Statistical analyses employed Wilcoxon signed-rank tests to assess tissue-specific SNR variations between reconstruction intensities (weak/medium/strong), given that all comparisons were performed on measurements from the same cohort of patients, paired statistical tests were mandatory. Comparisons between two related conditions were performed using the Wilcoxon signed-rank test. Pairwise comparisons were performed using Wilcoxon signed-rank tests α=0.05, specifically comparing the conventional reconstruction with each DLR level (Weak, Medium, Strong). Cohen's weighted kappa quantified interobserver agreement (two radiologists) and intra-observer concordance (radiologist vs. arthroscopic reference standard). The calculation formula is as follows: κ=Po−Pe/1−Pe, where Po is the observed proportion of agreement (i.e., the relative observed agreement among readers), Pe is the expected proportion of agreement by chance (23). Knee arthroscopy served as the diagnostic reference standard, with kappa analysis comparing DLR and non-DLR diagnostic accuracy. The kappa values were interpreted as: <0.20, poor; 0.21–0.40, fair; 0.41–0.60, moderate; 0.61–0.80, substantial; 0.81–1.00, near-perfect agreement. Statistical significance was defined as p < 0.05.

Wilcoxon signed-rank tests assessed SNR differences between conventional and DLR images across intensity levels (weak/medium/strong). Compared to conventional images, the DLR ones demonstrated significant SNR improvements (12.61–350.63%) across all tissue types evaluated in this study (p < 0.001), as detailed in Table 2 and Figure 2. Tissue-specific SNR increases ranged: Cartilage 17.23–54.16%; Meniscus 13.73–63.27%; Osseous 13.47–54.8%; Ligament 12.61–56.95%; Musculature 100.31–350.63%, with the most pronounced improvement in musculature. Moreover, the DLR performance exhibited sequence dependency: T2WI and PDWI demonstrated enhanced DLR sensitivity vs. T1WI, with quantifiably superior SNR improvements.

Interobserver agreement was quantified between two radiologists using Cohen's weighted κ coefficient. Table 3 provides the complete results of the image quality analysis. The interreader agreement was substantial for all quality measures (κ = 0.72–0.82). The two readers consistently scored the overall image quality as good or very good [rating, ≥4 (IQR, 4–5 or higher)]. The artifacts with DLRs were minimal or nearly artifact-free [rating, ≥4 (IQR, 4–5 or higher)] and the artifacts without DLR was moderate [rating, 4 (IQR, 4–4)]. On images processed with DLR, all evaluated anatomical structures, including cartilage, meniscus, bone, ligament, and muscule—were good or very good [rating, ≥4 (IQR, 4–5 or higher)]. At the strong reconstruction intensity level, these structures consistently attained the maximum score of 5. By comparison, conventional methods yielded scores of 3 or above for all anatomical structures. Figure 3 illustrates progressively enhanced trabecular bone visualization correlating with reconstruction intensity (arrows), while Figure 4 demonstrates superior lesion margin definition. Figure 5 demonstrates progressive noise reduction in fat-suppressed sequences with increasing reconstruction intensities, validating image quality enhancements while preserving acquisition durations and diagnostic fidelity.

Strong-intensity DLR demonstrated superior diagnostic concordance vs. non-DLR (κ = 0.908–1.00 vs. 0.882–0.963) when validated against arthroscopy-integrated radiologist assessments (Table 4). Strong-intensity DLR demonstrated perfect concordance with arthroscopy (κ = 1.00) vs. non-DLR (κ = 0.963) in patellar cartilage assessments. Strong-intensity DLR achieved perfect diagnostic concordance (κ = 1.00) for lateral meniscus and osseous structures, with lesion detection discrepancies ≤3% (e.g., 53/16 vs. 52/17). Muscle: Strong-intensity DLR demonstrated equivalent diagnostic accuracy vs. non-DLR (κ = 1.00) alongside enhanced noise reduction. Strong-intensity DLR demonstrated modest ACL/PCL concordance improvements (κ = 0.942/0.915 vs. 0.908/0.866) with reduced diagnostic ambiguity in ligament evaluations. Medial Meniscus: Strong-intensity DLR demonstrated enhanced diagnostic concordance vs. non-DLR (κ = 0.969 vs. 0.909), though subtle artifacts persisted (3/66 vs. arthroscopy 2/67). Femoral Cartilage: DLR reconstruction modes demonstrated equivalent diagnostic concordance (κ = 1.00) in structurally homogeneous lesions, suggesting limited incremental clinical value. Bone: Strong-intensity DLR and non-DLR demonstrated comparable diagnostic concordance (κ = 1.00 vs. 0.958), potentially attributable to intrinsic tissue contrast characteristics. Strong-intensity DLR enhances diagnostic reliability in noise-vulnerable anatomical zones, while requiring optimization for heterogenous tissue interfaces and artifact-susceptible regions.