Table 3.

Performances of the five models optimized for detecting of radiographic hip osteoarthritis on the CT-AP images within the validation and test datasets. The CT-AP images were created by sequentially summing the CT slices. Model-1 was trained with unprocessed X-ray images. Model-2 was trained similarly with downsampled X-rays to resemble CT-AP images. Model-3 and Model-4 were trained on a combination of the CT-AP and X-ray images where the X-ray images were similar to the ones used in Model-1 and Model-2, respectively. Model-5 was trained solely on the CT-AP images

Trained models	Accuracy	Balanced accuracy	Precision	Recall	F1-score	PR AUC [95% CI]	ROC AUC [95% CI]
CT-AP images of the validation dataset
Model-3	79.2	79.0	0.79	0.78	0.78	0.90 [0.79–0.94]	0.93 [0.73–0.98]
Model-4	87.5	83.3	0.83	0.92	0.87	0.88 [0.76–0.94]	0.91 [0.69–0.99]
Model-5	83.3	83.3	0.83	0.82	0.83	0.76 [0.52–0.91]	0.93 [0.71–0.99]
CT-AP images of the test dataset
Model-1	50.0	60.0	0.60	0.71	0.65	0.79 [0.63–0.93]	0.73 [0.49–0.89]
Model-2	50.0	60.0	0.60	0.71	0.65	0.72 [0.55–0.89]	0.63 [0.37–0.84]
Model-3	83.3	82.2	0.82	0.82	0.82	0.89 [0.79–0.94]	0.93 [0.72–0.99]
Model-4	75.0	80.0	0.80	0.80	0.80	0.87 [0.77–0.94]	0.87 [0.64–0.97]
Model-5	83.3	82.2	0.82	0.82	0.82	0.80 [0.57–0.92]	0.89 [0.67–0.97]

PR AUC, area under the precision recall curve; ROC AUC, area under the receiver operating characteristics curve; CI, confidence interval