A deep learning model for classification of chondroid tumors on CT images

BMC Cancer

Table 2 Performance parameters of the four readers and the deep learning algorithm for the differentiation between enchondromas vs. chondrosarcomas (G1-G3), enchondromas vs. acts as well as acts vs. High-Grade chondrosarcomas on the external dataset

Enchondromas vs. Chondrosarcomas (G1 - G3)
	Resident 1	Resident 2	Fellowship	Expert	Algorithm
Accuracy	0.770	0.840	0.870	0.900	0.750
Positive Predictive Value	0.772	0.847	0.844	0.889	0.736
Negative Predictive Value	0.762	0.821	0.957	0.929	0.846
Sensitivity	0.924	0.924	0.985	0.970	0.970
Specificity	0.471	0.676	0.647	0.765	0.324
Balanced Accuracy	0.697	0.800	0.816	0.867	0.647
F1-Score	0.841	0.884	0.909	0.928	0.837
AUC	0.810	0.891	0.910	0.932	0.884

ACT vs. High-Grad Chondrosarcoma
	Resident 1	Resident 2	Fellowship	Expert	Algorithm
Accuracy	0.545	0.545	0.727	0.652	0.561
Positive Predictive Value	0.686	0.686	0.719	0.685	0.719
Negative Predictive Value	0.387	0.387	0.778	0.500	0.412
Sensitivity	0.558	0.558	0.953	0.860	0.535
Specificity	0.522	0.522	0.304	0.261	0.609
Balanced Accuracy	0.540	0.540	0.629	0.561	0.572
F1-Score	0.615	0.615	0.820	0.763	0.613
AUC	0.525	0.541	0.630	0.561	0.637

Reader 1 and 2 = residents; Reader 3 = fellowship trained radiologist; Reader 4 = musculoskeletal attending radiologist; AUC = area under the receiver operating characteristic. curve

ISSN: 1471-2407