Table 1:

Average AUC and Kappa scores for rotational (R) and translational (T) instabilities. All columns are compared to the patient-level expert consensus Tile grade by three radiologists. $\mathbf{B}\mathbf{M}\left(GT\right)$ represents Bayesian model (BM) fitness when using the reference standard fracture detections used to train the Faster-RCNN model; $\mathbf{B}\mathbf{M}\left(F{X}_{low}\right)$ represents BM prediction on automatically detected fractures when lower confidence fractures (with $z=0.5$) are immediately included in Tile grade inference; $\mathbf{B}\mathbf{M}\left(F{X}_{high}\right)$ represents BM prediction on automatically detected fractures with high confidence $(z=max)$; BM refinement represents predictions after the proposed refinement pipeline.

	$\mathrm{BM}\left(GT\right)$		$\mathrm{B}\mathrm{M}\left(F{X}_{low}\right)$		$\mathrm{B}\mathrm{M}\left(F{X}_{high}\right)$		BM refinement
Metrics	R	T	R	T	R	T	R	T
AUC	0.81	0.84	0.71	0.63	0.78	0.82	0.85	0.83
Kappa	0.32	0.48	0.1	0.18	0.15	0.38	0.24	0.5