Automatic segmentation of fat metaplasia on sacroiliac joint MRI using deep learning

Insights into Imaging

Table 3 Performance of the segmentation deep learning model, radiologists without and with model assistance on the external test set

	DSC (%)	Precision (%)	Recall (%)
2.5D-AttentionUNet model	85.44 ± 6.09	85.83 (82.62–89.04)	86.43 (81.10–91.76)
Radiologists
Radiological resident	75.70 ± 10.87	66.18 (59.69–72.68)	91.13 (87.71–94.55)
p value^a	0.001	/	/
Expert radiologist	85.03 ± 9.72	80.32 (74.71–85.92)	91.11 (86.84–95.38)
p value^a	0.874	/	/
Model-assisted radiologists
Radiological resident	82.87 ± 6.88	76.18 (72.01–80.35)	92.05 (88.14–95.95)
p value^b	< 0.001	/	/
Expert radiologist	85.74 ± 8.08	81.59 (76.84–86.33)	91.39 (87.45–95.33)
p value^b	0.496	/	/

DSC is presented as an average percentage with a standard deviation. Precision and recall are shown as percentages with 95% confidential intervals. p values less than 0.05 show statistical differences
^aData were compared between 2.5D-AttentionUNet model and radiologists
^bData were compared between radiologists and model-assisted radiologists. DSC Dice similarity coefficient