Synopsis

Keywords: Image Reconstruction, Cartilage

Diagnostic performance was limited to image resolution and contrast between target tissues and surrounding tissues. A rapid knee imaging has been perused but no loss of image quality is critical. This study proposed a rapid knee imaging based on two-dimensional fast spin echo sequence and examined the reliability and diagnostic performance of deep learning-based reconstruction T1-, T2- and PD- weighted images on knee joint pathology via comparison of images with and without deep-learning reconstruction algorithm (DLR). Diagnostic efficacy on knee structural abnormalities of 2D DLR FSE sequence elevated using knee arthroscopy results as the gold standard.

Introduction and Purpose

Reference standards for knee MRI are proton density (PD)– and T1–weighted fast spin-echo (FSE) sequences due to the excellent tissue contrast and high in-plane spatial resolution with good assessment of meniscal, ligamentous, and cartilaginous injuries. Limited to scan time, images often have lower resolution and poor image quality. Early detection of knee osteoarthritis is difficult because the contrast between the articular cartilage and the surrounding tissues is low. Deep learning reconstruction (DLR) algorithm has showed benefits for clinical imaging field for better image quality, less noise, and shorter scan time[1]. Our study aimed to propose a rapid deep learning reconstruction-based knee imaging protocol and ensured its feasibility in clinical knee imaging via evaluation of the image quality and diagnostic performance of 2D fast spin echo (FSE) sequences on knee.

Materials and methods

With approval of the Institutional Review Board, a total of 92 patients in supine position underwent knee routine MRI with a protocol including 2D FSE-based T1-, T2-, and PD-weighted images using 18-channel knee coils (Table 1). Among 92 patients, the arthroscopic findings of 22 patients in this prospective study were used as the reference standard for diagnosis. Two datasets of original images (FSE_O) and DLR images (FSE_DL) were automatically generated. Diagnostic performance and image quality of FSE_O and FSE_DL was independently evaluated by radiologists with 3 to 9 years of experience in interpreting musculoskeletal MRI, according to International Cartilage Regeneration&Joint Preservation Society (ICRS) [2]and the 5-point Likert scale. Objective evaluation of image quality such as SNR and CNR were also obtained. Assessment of pathologies and internal derangement were conducted by the same two radiologists and included the evaluation of the medial and lateral menisci; anterior and posterior cruciate ligaments; and cartilage defects of the medial and lateral femur trochlea, the medial tibia plateau, the trochlear groove, and the retropatellar cartilage. Structural abnormal-ities were graded as 0 = normal, 1 = altered (degenerative, postoperative), and 2 = tear. Areas of bone marrow edema (femoral, patellar, tibial), as well as fractures and joint effusion, were evaluated being present or absent. If there were discrepancies between the readers, a consensus reading was enclosed to define false-positive and false-negative findings. Evaluation was performed repeatedly with an interval of at least 2 weeks. Assessment was recorded and analyzed, in particular the correlation with arthroscopic findings using SPSS (version 25.0, IBM Corp). P<0.05 was considered statistical difference.

Results

The overall image quality, sharpness and diagnostic confidence for FSE_DL were higher compared to FSE_O, showing significantly improved sharpness (p < 0.05). Inter- and intra-reader agreement was substantial to almost perfect (ICC=0.710 -0.898) (Table 2). In objective evaluation, SNR and CNR of PDWIDL and T1WI_DL images were significantly higher than that of PDWI_O and T1WI_O images (p < 0.05) (Table 3). Two radiologists assessed the sequences regarding structural abnormalities of the knee based on FSE_O and FSE_DL (Table 4). Inter- and intra-reader agreement were moderate-excellent consistent (κ = 0.792-1.000) for the detection of internal derangement. Intra-reader agreement was substantial to almost perfect (κ=0.769-0.771) for the assessment of cartilage defects and almost perfect (κ = 0.944-1.000) for the assessment of meniscal, ligament. When we classified the reference diagnoses as normal and abnormal, the sensitivities of FSE_DL ranged from 92.3% to 100% compared with 66.7% to 100% for FSE_O, and the specificities of both FSE_DL and FSE_O ranged from88.9% to 100%. The accuracy of FSE_DL ranged from 83.3% to 100% compared with 81.8 to 100% for FSE_O. There were no significant differences in sensitivity, specificity and accuracy between FSE_DL and FSE_O for diagnostic performance (P = 0.309–1.000). A total of 10 cartilage lesions were detected by reader 1 on FSE_DL compared with 8 lesions were detected on FSE_O,and 11 lesions were detected on FSE_DL compared with 10 lesions were detected on FSE_O by reader 2. The diagnostic accuracy of cartilage injury on FSE_DL is higher than FSE_O. However for two readers, there was no statistically significant difference between the two images for the detection of these structural abnormalities. (all, P > 0.05).

Discussion and conclusions

In this study, deep learning technology is embedded in the stage of MRI reconstruction of original data, and then the noise of MRI is effectively removed in the stage of original data acquisition, and finally pure MR Signal is achieved. Therefore, high signal-to-noise ratio and high-resolution images can be achieved at faster scanning speed. In terms of cartilage defect classification, the diagnostic accuracy of cartilage injury on FSE_DL is higher than FSE_O. This may attribute to DL-based image reconstruction, leading to improved image quality and increased contrast between cartilage and neighboring structure. The lesion clarity elevated. 2D FSE sequences of the knee using deep-learning reconstruction are clinically feasible, showing excellent image quality and improving diagnostic efficacy compared to the original images.

Acknowledgements

The author would like to thank Guangnan Quan for technical assistance.