TABLE III.

The Performance Scores of the Multi‐Modal Deep Learning Model With Different Fusion Strategies and an Experienced Radiologist.

Training Input		Vector	Accuracy	ROC‐AUC	Sensitivity	Specificity
T1 T2 CE‐T1	Early‐fusion		0.83 (0.73–0.92)	0.88 (0.74–0.98)	0.89 (0.67–1)	0.82 (0.70–0.93)
	Late‐fusion	Add.	0.87 (0.78–0.96)	0.88 (0.75–1)	0.69 (0.33–1)	0.91 (0.80–1)
		Spl.	0.85 (0.76–0.94)	0.89 (0.72–0.99)	0.70 (0.44–1)	0.89 (0.78–0.98)
T2 CE‐T1 DWI‐b1000	Early‐fusion		0.76 (0.63–0.88)	0.86 (0.74–0.96)	0.70 (0.44–1)	0.77 (0.63–0.90)
	Late‐fusion	Add.	0.85 (0.73–0.94)	0.93 (0.83–0.99)	0.80 (0.56–1)	0.86 (0.75–0.95)
		Spl.	0.85 (0.75–0.93)	0.96 (0.89–1)	0.90 (0.67–1)	0.84 (0.73–0.95)
DWI‐b0 DWI‐b1000 ADC	Early‐fusion		0.82 (0.69–92)	0.83 (0.65–0.96)	0.60 (0.22–0.89)	0.86 (0.75–0.98)
	Late‐fusion	Add.	0.83 (0.71–0.94)	0.87 (0.67–1)	0.79 (0.56–1)	0.84 (0.73–0.95)
		Spl.	0.94 (0.88–0.1)	0.91 (0.7–1)	0.89 (0.67–1)	0.95 (0.88–1)
All MRI sequences	(Early + late)‐fusion	Add.	0.78 (0.65–0.88)	0.76 (0.56–0.91)	0.59 (0.33–0.89)	0.82 (0.70–0.93)
		Spl.	0.76 (0.63–0.88)	0.84 (0.70–0.96)	0.69 (0.44–1)	0.77 (0.65–0.90)
	Late‐fusion	Add.	0.89 (0.80–0.96)	0.80 (0.56–1)	0.70 (0.33–1)	0.93 (0.85–1)
		Spl.	0.81 (0.69–0.92)	0.95 (0.86–1)	0.81 (0.56–1)	0.81 (0.70–0.93)
Radiologist			0.70 (0.58–0.82)	0.74 (0.59–0.86)	0.70 (0.44–1)	0.70 (0.58–0.85)

Values in the parentheses indicate 95% confidence interval. Notable values are in bold, see main text for details. “Add.” = vector addition method, “Spl.” = vector splicing method.

ADC = apparent diffusion coefficient; DWI = diffusion‐weighted; MRI = magnetic resonance image; ROC = receiver operator characteristic.