Predictive performance of radiomic models based on features extracted from pretrained deep networks

Insights into Imaging

Table 4 AUCs of the best-performing models for each dataset

Dataset	Parameters of deep model					AUC (deep features)	Parameter of generic model
Dataset	Architecture	Extraction level	Slices	Aggregation	Segmentation	AUC (deep features)	Discretization	AUC (generic features)	P (deep vs generic)	AUC (from other studies)
C4KC-KiTS	ResNet-50	Mid	All	Max	ROIchannel	0.75	binWidth:25	0.76	− 0.01 (p = 0.716)	0.88 [18]
CRLM	ResNet-18	Top	Max	Mean	ROIchannel	0.81	binWidth:10	0.79	0.02 (p = 0.715)	0.68 (0.56–0.8) [19]
Desmoid	ResNet-50	Mid	All	Max	ROI	0.86	binWidth:10	0.89	− 0.03 (p = 0.456)	0.82 (0.75–0.89) [19]
GIST	VGG-19	Mid	Max	Mean	ROIchannel	0.79	binWidth:25	0.78	0.01 (p = 0.674)	0.77 (0.71–0.83) [19]
HN	VGG-19	Mid	All	Mean	ROIchannel	0.91	binCount:50	0.89	0.02 (p = 0.405)	0.84 (0.77–0.91) [19]
ISPY-1	VGG-19	Mid	All	Mean	ROIchannel	0.78	binWidth:25	0.7	0.08 (p = 0.109)	–
Lipo	ResNet-18	Top	All	Mean	ROIcut	0.92	binWidth:25	0.9	0.02 (p = 0.535)	0.83 (0.75–0.91) [19]
Liver	DenseNet169	Top	All	Max	ROIcut	0.82	binWidth:100	0.79	0.03 (p = 0.526)	0.81 (0.75–0.87) [19]
Melanoma	DenseNet169	Mid	All	Mean	ROIcut	0.82	binCount:50	0.72	0.1 (p = 0.176)	0.51 (0.4–0.62) [19]
TCGA-GBM	ResNet-50	Mid	Max	Mean	ROIcut	0.89	binWidth:25	0.79	0.1 (p = 0.176)	–

AUC of the best-performing models for each dataset. Statistical difference was tested with a DeLong test. AUCs (with 95% CI where applicable) from other studies were reported in the last column; for ISPY-1 and TCGA-GBM, no corresponding studies using only a single MR-weighting were found