Approval for Research in Human Subjects
This study was approved by the institutional review board of Hospital for Special Surgery and conducted in accordance with the Health Insurance Portability and Accountability Act. Informed consent was obtained from all individual participants.
Study Design and Study Population
In total, 29 subjects were prospectively enrolled and in these subjects 60 peripheral nerves were evaluated. These nerves were chosen for evaluation as they are commonly evaluated in clinical practice and were the largest in diameter within the imaged field-of-view. Written informed consent was obtained from all patients prior to imaging performed between February 2019 and November 2019.
Inclusion Criteria
All patients who presented to our institution for standard-of-care MR neurography evaluation for clinically suspected neuropathy were considered for study inclusion.
Exclusion Criteria
Exclusion criteria were standard MRI safety contraindications. No patients were excluded from the study.
Image Acquisition
All scans were performed on a 3.0 Tesla clinical scanner (MR750, GE Healthcare) using a 16-channel flexible coil (Neocoil). Axial, 2D intermediate-weighted fast spin echo sequences (FSE) were obtained as part of the institution’s standard MR neurography protocol, using the following acquisition parameters optimized for each body part: echo time (TE): 21-39 ms, repetition time (TR): 3177-6522 ms, receiver bandwidth (RBW): 195.312-488.281 Hz/pixel, field of view: 80-320 mm, acquisition matrix: (256 or 512)x(256-512), slice thickness: 2.0-4.5 mm, echo train length (ETL): 8-21, number of excitations (NEX): 1-4, parallel imaging factors between 1.5 and 2. No compressed-sensing acquisition was utilized.
Image Reconstruction
In addition to the untouched SOC reconstruction (i.e. images immediately derived from the scanner), the raw data was retrospectively reconstructed with a deep convolution neural network, i.e. “DLRecon”, [11] a vendor-provided software installed on the scanner. This neural network accepts raw unfiltered complex valued image inputs and outputs images with higher SNR and reduced truncation artifacts by utilizing a feed-forward approach.[11] Gibbs ringing that occurs near sharp edges is also removed by the neural network, resulting in increased image sharpness.[11]
DLRecon was previously trained using pairs of images containing conventional MR images and ‘near-perfect’ images, defined as those with high resolution, minimal ringing, and very low noise levels.[11] Four million unique image pairs were employed in this supervised learning approach, and image augmentations such as rotations, flips, intensity gradients, phase manipulations, and Gaussian noise were used to increase the robustness of the training set.[11] The training images were diverse, allowing generalizabitlity of DLRecon’s application across anatomical locations.[11] Additionally, the network was trained using a gradient backpropogation and ADAM optimizer.[11,15]
Image Analysis
Anonymized images were evaluated on a picture archiving and communication system (PACS) (Sectra V18.1, Sectra AB) by two board-certified radiologists: reader 1 (DBS) with 6 years of dedicated MR neurography experience, and reader 2 (AJB) with 10 years of general musculoskeletal MRI experience. Both readers underwent a training session to establish grading consensus by reviewing both SOC- and DLRecon-MRI images from 10 separate datasets not included in the analysis. Study images were randomized prior to evaluation by the 2 readers, who remained blinded with respect to the post-processing method. Each reader independently scored images for outer epineurium conspicuity and visualization of fascicular architecture by reviewing the entire volume for slices that best demonstrated the nerve in an orthogonal plane, as this is the best plane to visualize fascicular architecture. For evaluation of pulsation artifact, ghosting artifact, and bulk motion, readers considered all slices. Outer epineurium conspicuity was defined by the number of distinct borders visualized between the nerve and immediately surrounding perineural fat (maximum 4: anterior, posterior, medial, and lateral). Fascicular architecture was graded using the following Likert scale: 1-poor visualization, 2-average visualization, 3-good visualization, 4-excellent visualization. The presence of pulsation artifact, ghosting artifact, and bulk motion was graded using the following scale: 0-none, 1-mild, 2-moderate, 3-severe. Additionally, each radiologist was asked to ‘guess’ as to whether each image dataset was processed with SOC-MRI or DLRecon-MRI to determine the extent of potential bias from perceiving image texture differences.
Statistical Analysis
Statistical analyses were performed by a biostatistician (BL) with 5 years of experience. Odds ratios (OR) and 95% confidence intervals (CI) obtained from marginal ordinal logistic regression models estimated with generalized estimating equation were used to evaluate for differences in grades between DLRecon- and SOC- MRIs. Patients were treated as the repeated factor to account for any within-patient correlations between image type as well as patients from whom more than one nerve was evaluated in their exam. Given that ORs were calculated as a comparison of DLRecon-MR and SOC-MR images, an OR of 1 was interpreted as no difference between DLRecon- and SOC-MRI; an OR >1 was interpreted as the DLRecon-MRI being more likely to have a higher grade than the SOC-MRI; and an OR <1 was interpreted as the SOC-MRI being more likely to have a higher grade than the DLRecon-MRI. Statistical significance was set a priori to p< 0.05.
Agreement between DLRecon- and SOC-MRI grades for each reader (inter-reconstruction agreement) was analyzed using ordinal-weighted Gwet’s agreement coefficients (AC). Clustered bootstrap confidence intervals were used to account for patients who had more than 1 nerve examined. Strength of agreement was determined using the following scale: <0 = poor, 0.00-0.2 = slight, 0.21-0.40 = fair, 0.41-0.60 = moderate, 0.61-0.80 = substantial, and 0.81-1 = almost-perfect.[16] The interrater agreement was analyzed in the same manner. Statistical analyses were performed with SAS v. 9.4 (SAS Institute).