Prediction of lipomatous soft tissue malignancy on MRI: comparison between machine learning applied to radiomics and deep learning

European Radiology Experimental

Table 2 Sensitivity, specificity, AUC (mean ± standard deviation) obtained for each dataset (images, radiomics with and without batch effect normalisation) and classifier combination over cross-validation obtained using 10 cross-validation folds on the test cohort

Dataset	Classifier	Sensitivity (%)	Specificity (%)	AUC
Images	CNN	90 ± 0.21	10 ± 0.10	0.53 ± 0.09
	ResNet50	81.9 ± 0.06	57.7 ± 0.07	0.8 ± 0.11
	FE + XGB	80.3 ± 0.12	57 ± 0.11	0.78 ± 0.13
Radiomics	LR	77 ± 0.19	62.5 ± 0.24	0.84 ± 0.12
	SVM	75 ± 0.14	62.5 ± 0.26	0.83 ± 0.43
	RF	84 ± 0.01	62.5 ± 0.12	0.79 ± 0.65
	GB	72.5 ± 0.05	72.5 ± 0.16	0.83 ± 0.12
Radiomics with batch correction	LR	70 ± 0.22	77.5 ± 0.31	0.86 ± 0.13
	SVM	75 ± 0.17	70 ± 0.25	0.82 ± 0.15
	RF	100 ± 0.18	92.5 ± 0.33	0.96 ± 0.04
	GB	98 ± 0.20	87.5 ± 0.13	0.99 ± 0.02

AUC area under the curve, CNN convolutional neural networks, FE feature extraction, XGB Xgboost, GB gradient boosting, LR logistic regression, RF random forest, SVM support vector machine