Table 2.

Summary of AI studies for fully automated meniscus tear detection

						Diagnostic performance of AI algorithm			Diagnostic performance of human readers
Study	Reference standard	Label	Analyzed sequence	Field strengths [T]		Both	Med	Lat	Both	Med	Lat	Comments
Bien et al. [11]	Radiologist interpretation	Intact, tear	Sag T2, cor T1, ax PD	1.5, 3.0	Sensitivity Specificity AUC	71% 74% 85%	-	-	82% 88% -	-	-
Pedoia et al. [31]	Radiologist interpretation	Intact, tear	3D fat-suppressed PD	3.0	Sensitivity Specificity AUC	82% 90% 89%	-	-	-	-	-	Severity grading into no tear, mild-moderate tear, and severe tear achieved accuracies of 81%, 78%, and 75%, respectively
Fritz et al. [32]	Surgical inspection	Intact, tear	Sag and cor fat-suppressed fluid-sensitive	1.5, 3.0	Sensitivity Specificity AUC	91% 87% 96%	84% 88% 88%	58% 92% 78%	94% 90% 92%	95% 88% 92%	69% 97% 83%	Performances of human readers are given as averages
Rizk et al. [33]	Radiologist interpretation	Intact, tear	Cor fat-suppressed PD, sag fat-suppressed PD	1.0, 1.5, 3.0	Sensitivity Specificity AUC	-	89% 84% 93%	67% 88% 84%	-	-	-	Validation with external “MRNet” data set [11] achieved a sensitivity, specificity, and AUC of 77%, 84%, and 0.83, respectively
Irmakci et al. [17]	Radiologist interpretation	Intact, tear	Sag T2, cor T1, ax PD	1.5, 3.0	Sensitivity Specificity AUC	62–69% 76–81% 78–81%	-	-	-	-	-
Tsai et al. [15]	Radiologist interpretation	Intact, tear	Cor T1	1.5, 3.0	Sensitivity Specificity AUC	86% 89% 90%	-	-	-	-	-

Sag sagittal, cor coronal, ax axial, PD proton density, TSE turbo spin echo, AUC area under the receiver operating curve, 3D three-dimensional, AI artificial intelligence, both both menisci combined, med medial meniscus, lat lateral meniscus