Hierarchical Severity Staging of Anterior Cruciate Ligament Injuries\n using Deep Learning with MRI Images

Abstract

Purpose: To evaluate the diagnostic utility of two convolutional neural\nnetworks (CNNs) for severity staging of anterior cruciate ligament (ACL)\ninjuries.\n Materials and Methods: This retrospective analysis was conducted on 1243 knee\nMR images (1008 intact, 18 partially torn, 77 fully torn, and 140 reconstructed\nACLs) from 224 patients (age 47 +/- 14 years, 54% women) acquired between 2011\nand 2014. The radiologists used a modified scoring metric. To classify ACL\ninjuries with deep learning, two types of CNNs were used, one with\nthree-dimensional (3D) and the other with two-dimensional (2D) convolutional\nkernels. Performance metrics included sensitivity, specificity, weighted\nCohen's kappa, and overall accuracy, followed by McNemar's test to compare the\nCNNs performance.\n Results: The overall accuracy and weighted Cohen's kappa reported for ACL\ninjury classification were higher using the 2D CNN (accuracy: 92% (233/254) and\nkappa: 0.83) than the 3D CNN (accuracy: 89% (225/254) and kappa: 0.83) (P =\n.27). The 2D CNN and 3D CNN performed similarly in classifying intact ACLs (2D\nCNN: 93% (188/203) sensitivity and 90% (46/51) specificity; 3D CNN: 89%\n(180/203) sensitivity and 88% (45/51) specificity). Classification of full\ntears by both networks were also comparable (2D CNN: 82% (14/17) sensitivity\nand 94% (222/237) specificity; 3D CNN: 76% (13/17) sensitivity and 100%\n(236/237) specificity). The 2D CNN classified all reconstructed ACLs correctly.\n Conclusion: 2D and 3D CNNs applied to ACL lesion classification had high\nsensitivity and specificity, suggesting that these networks could be used to\nhelp grade ACL injuries by non-experts.\n

Ähnliche Arbeiten