A transfer learning approach for staging diagnosis of anterior cruciate ligament injury on a new modified MR dual precision positioning of thin-slice oblique sagittal FS-PDWI sequence.

Abstract

To evaluate the diagnostic performance of the transfer learning approach for grading diagnosis of ACL injury on a new modified dual precision positioning of thin-slice oblique sagittal FS-PDWI (DPP-TSO-Sag-FS-PDWI) sequence. And compare the prediction performances between artificial intelligence (AI) and radiologists. Patients with both DPP-TSO-Sag-FS-PDWI sequence and arthroscopic results were included. We performed a transfer learning approach using the pre-trained EfficientNet-B0 model, including whole image and regions of interest (ROI) image inputs, and reset its parameters to achieve an automatic hierarchical diagnosis of ACL. A total of 235 patients (145 men and 90 women, 37.91 ± 14.77years) with 665 images were analyzed. The consistencies of AI and arthroscopy (Kappa value > 0.94), radiologists and arthroscopy (Kappa value > 0.83, p = 0.000) were almost perfect. No statistical difference exists between the whole image and radiologists in the diagnosis of normal ACL (p = 0.063) and grade 3 injury (p = 1.000), while the whole image was better than radiologists in grade 1 (p = 0.012) and grade 2 injury (p = 0.003). The transfer learning approach exhibits its feasibility in the diagnosis of ACL injury based on the new modified MR DPP-TSO-Sag-FS-PDWI sequence, suggesting that it can help radiologists hierarchical diagnose ACL injuries, especially grade 2 injury.